betaFeedback

WorksComputer Science

EviScope: Evidence-Conditioned Sampling for Long-Range 3D Detection

Abstract

Query-based camera-radar detectors choose where and over which frames each object query samples from the query's position or feature. We test whether radar evidence around the query should also set these choices at the 150 m range of MAN TruckScenes, where recall differs across object-level radar-return counts within every range band. EviScope is a minimal test: the current frame's returns form a non-learned bird's-eye evidence field, which each query reads at its decoded centre to scale its radial sampling extent and to weight eight frames. It adds 5,450 parameters (0.009%) to RaCFormer, with no latency increase in our timing. Our module-free recipe exceeds the published TruckScenes mAP (25.9 over seven runs against 24.1), so we measure the module only against its own baseline at matched seeds. The evidence a query reads rises with its object's return count (Spearman 0.51) but captures zero-return objects only in part, and the learned radius tracks it (ρ≈−0.91ρ\approx-0.91 in four runs). Over seven four-epoch pairs the module changes NDS by +0.31±0.36+0.31\pm0.36 and mAP by −0.27±0.22-0.27\pm0.22 (mean ±\pm SE). Against a geometry-only conditioner with the same parameter count, trained at three matched seeds, it is higher by +1.25±0.72+1.25\pm0.72 NDS and +0.87±0.21+0.87\pm0.21 mAP, on mAP in every pair, and the control's mean NDS is near the baseline's. An exploratory far-range gain failed an internally pre-specified held-out test, and large inference-time substitution penalties also occur without query-local evidence: neither learned allocation nor inference-time dependence substitutes for matched end-to-end comparisons.

Introduction

Camera–radar fusion combines the semantic detail of vision with radar’s range and velocity measurements for long-range 3D object detection [7, 33, 19, 25]. Detectors of this kind are query-based: they refine a fixed set of object queries over successive decoder layers and sample image and bird’s-eye-view (BEV) features from each query’s position [4, 45, 27, 26]. Recent work exploits geometric structure through concentric query initialisation with adaptive sampling [5], distance-modulated radar aggregation [46], or object-centred Gaussian representations [1]. We investigate whether explicit local radar evidence provides conditioning information for query sampling beyond query geometry.

Our analysis of the 86,252 evaluated objects of the MAN TruckScenes validation split [7] reveals substantial recall differences across radar-return counts within every range band (Figure 1a). Baseline recall increases from 71.4% for objects with no radar return to 89.6% for those with more than fifteen. Range alone therefore does not fully characterise detection difficulty, which motivates local radar evidence as an additional sampling cue.

Figure 1. (a) Recall of our baseline detector on TruckScenes val by number of radar returns on the object, one curve per range band (Table 20). Recall differs across return-count bins within each band, but is not monotonic in return count; zero-return recall is also not monotonic in range. (b) NDS of the same detector with and without EviScope, seven matched seeds, with the paired difference above each pair. The NDS axis starts at 35.

We propose EviSCOPE, an evidence-conditioned query sampling framework that adapts the spatial sampling extent and the temporal aggregation weights while keeping the query position as the geometric reference. Current-frame radar returns form a parameter-free BEV evidence field of multiscale return counts and radar measurement statistics. Each query reads this field at its decoded centre and combines the descriptor with its own geometry through a lightweight shared network. A spatial branch and a temporal branch then predict a radial sampling scale and a weighting over the eight fused frames, so queries at similar positions but with different evidence sample differently; a training-only density branch supervises the shared representation. The change is small: one non-learned input conditions two sampling controls that RaCFormer already has. Matched-capacity controls test its contribution; Table 15 separates inherited and added components. It adds 5,450 parameters, 0.009% of the detector, with no latency increase in our timing protocol.

Our experiments separate the performance of the training configuration from the module’s incremental effect. At eight epochs the baseline and EviSCOPE reach 26.4 / 39.2 and 26.7 / 39.9 mAP / NDS, both above the strongest published entry in our comparison (24.1 / 38.4), using a smaller backbone and lower resolution; at the 256 × 864 input of the published ResNet-50 entry, one EviSCOPE run reaches 28.2 / 39.9 against its 23.6 / 37.4. These are not comparisons at matched compute (§ 5.2). Across seven seed-matched pairs at four epochs the module’s paired effect is +0.31 ± 0.36 NDS and −0.27 ± 0.22 mAP (mean ± SE; Figure 1b). Against a geometry-only conditioner with the same parameter count, at three matched seeds, the module is higher by +1.25 ± 0.72 NDS and +0.87 ± 0.21 mAP, and the control’s learned radius stays at its lower bound for four out of five query Appendices. On the seed-0 pair grouped recall rises in 16 of the 19 scored range-by-evidence cells, but the far-range gain tested under a criterion fixed after the first seeds did not replicate on the held-out pairs (§ 5.3–§ 5.4).

The emitted radial scale tracks local evidence (rank correlations −0.92 to −0.90 in four runs), and uniform temporal weights at inference cost 11.0 ± 2.0 NDS (mean ± SD), but an evidence-blind control also loses 11.75 (§ 6).

Contributions

  • A variance-characterised recipe and baseline for camera–radar detection on TruckScenes: seven ResNet-50 runs at four epochs average 25.9 mAP (SD 0.45) against 24.1 for the best published entry, with code in the supplementary material (§ 5.2).

  • EviSCOPE and a controlled test of evidence conditioning: the module, geometry-only and shuffled-evidence controls with the same parameter count, the density branch at three

matched seeds, two eight-epoch pairs, and the measured link between object-level and query-local evidence (§ 4.2, Appendices § 29 and § 30).

  • Four signals often read as evidence of value, namely an evidence-tracking allocation (ρ≈−0.91\rho\approx-0.91), 11-point substitution penalties, a +5-point subset gain and a +1.5 NDS two-seed average, all occur here without a replicated four-epoch end-to-end gain; matched seeds and held-out tests separate them (§ 5.3–§ 6).

Related Work

Long-range camera–radar 3D detection. Camera–radar fusion builds on the camera-only BEV lineage ([39, 14, 21, 27]) and is organised around object-level association, BEV fusion, and sparse cross-modal interaction. At the object level, Nabati & Qi [33] associate radar detections with image objects through a frustum and Kim et al. [18] exchange spatial and contextual cues between proposals with cross-attention. In BEV, Kim et al. [19] use radar to assist the view transform and absorb the residual misalignment with multi-modal deformable attention, while Lin et al. [25] align the branches with a radar-specific encoder and cross-modal attention, a design Lin et al. [24] extend to query-based detectors (see also Wolters et al. [47]). At truck range the cost of dense interaction binds: 4D-radar benchmarks ([36, 35]) annotate shorter ranges, and the entries that reach 150 m on TruckScenes (Fent et al. [7]) turn sparse: Wolters et al. [46] fuse inside sparse frusta with range-adaptive radar aggregation, and Bang et al. [1] build a sparse representation from primitives at estimated keypoints (see also Li et al. [20]). All of them allocate sampling by position (§ 1).

Modelling the radar return, and reliability-aware fusion. Radar returns are uncertain and the two modalities differ in observation quality (Huang et al. [15]). Long et al. [30] predict the distribution of returns an object would produce from its class, size and orientation and use it as a post-hoc matching kernel, which presumes at least one hit. Gu et al. [9] densify the points by modelling direction-of-arrival error. Yue et al. [49] expand radar points into Gaussians from cross-section and velocity priors and gate fusion on camera confidence; Park et al. [37] supervise a per-pixel reliability map from LiDAR–radar–image feature similarity. Both assume a dependable fallback modality. In our TruckScenes validation analysis, 42.7% of evaluated objects have no annotated radar return.

Query-driven adaptive spatio-temporal sampling. Query-based detectors use their observations through spatial sampling and aggregation across frames. Wang et al. [45] project a 3D reference into the multi-view images; Li et al. [22] fuse past BEV features with temporal self-attention; Lin et al. [23] propagate sparse instance features recurrently instead of re-sampling every past image. Liu et al. [26] generate sampling offsets in space and time from the query feature and mix the sampled features with predicted weights, and Chu et al. [5] combine polar initialisation with cross-view and cross-modal ray sampling. These policies condition on position or the query feature; we supply local radar statistics as an explicit, non-learned input to the spatial scale and the temporal weights, and test what that input adds beyond geometry (Appendix § 30).

Problem formulation

Recall by range and radar-return count

For every object the devkit evaluates on TruckScenes val we take the number of radar returns inside its box, which the dataset annotates, and check whether our baseline detector (§ 5.1) places any prediction within 2 m of its centre. We use recall rather than mAP because mAP on a subset of the ground truth is distorted by false positives that belong to other subsets.

Pooled recall differs across return-count bins. Table 20 in Appendix § 25 gives the numbers behind Figure 1a; its last row pools the ranges. 42.7% of the evaluated objects receive no return at all, and these are recovered at 71.4%, against 89.6% for objects with more than fifteen returns. The pooled relation is monotone across the five bins. On the full trainval annotation set, 774,688 boxes, the zero-return share varies by 57 points across range for cars and by 82 points across classes at 100–150 m, while the largest spread across any weather or time-of-day tag in a band is 10 points

range0 hits1–23–56–15>15
0–25 m82.0 (3,885)89.3 (4,042)93.4 (3,455)96.1 (4,451)90.2 (1,291)
25–50 m70.1 (9,950)83.3 (6,512)89.1 (3,412)88.1 (2,542)92.4 (1,341)
50–100 m68.9 (14,259)81.6 (8,688)80.8 (3,435)78.0 (2,682)83.8 (741)
100–150 m72.2 (8,758)80.9 (4,243)74.3 (1,643)70.7 (899)—
all ranges71.4 (36,852)83.3 (23,485)85.9 (11,945)87.5 (10,574)89.6 (3,396)

Table 20. Recall (%) of our baseline detector (seed 0) on TruckScenes val by range × radar-hit count, official 2 m matching, no score threshold; object counts in parentheses, cells below 30 objects omitted. Hit counts are the official num_radar_pts annotation. The last row pools the ranges over all 86,252 evaluated objects.

(Appendix L). These are descriptive associations, without adjustment for differences in class or scene composition. At 100–150 m, 72.6% of cars and 23.4% of trucks carry no return in the current frame.

Differences persist within range bands. In the nearest band, zero-return recall is 82.0%, versus 89.3–96.1% in the nonzero bins. However, recall is not monotonic in return count within any of the four bands. At 100–150 m, for example, recall in the 6–15 bin is 70.7%, below the zero-return value of 72.2%. These cross-tabulations motivate investigating radar evidence, but do not establish that it predicts detection better than range or adds predictive information after controlling for class, size and other object attributes.

The motivating statistic uses ground-truth boxes, whereas the module below reads a local field at a predicted query centre. The quantities can differ when a query is mislocalised or samples nearby clutter. Appendix S measures their correspondence, which is graded but partial; the object-level association motivates a hypothesis about query conditioning rather than validating that hypothesis.

Evidence-conditioned query sampling

We test whether current-frame local radar evidence can guide a query’s sampling extent and temporal aggregation. Sparse evidence motivates a wider search and a different weighting of past frames, but does not identify which past frames contain useful returns. We implemented this as a small addition to RaCFormer [5], shown in Figure 2. Everything outside the module, including the polar queries, ray sampling, the multi-frame samplers and the heads, is RaCFormer’s; Table 15 separates inherited and added components.

Figure 2. (a) The inherited RaCFormer detector, redrawn compactly; EVISCOPE attaches only to the ray-sampling step of the shared decoder layer. (b) The module. The current frame’s returns form a non-learned field EE; each query reads it at its decoded centre cqc_q (eqe_q) and joins its geometry g(q)g(q); a shared MLP feeds the radial extent αq\alpha_q, the frame weights wq=8πqw_q = 8\pi_q and a training-only density head (dashed). Sample locations and weights are schematic.

Evidence field. Radar returns for the current keyframe (the five most recent sweeps of each radar) are scattered into a 128×128128 \times128 bird’s-eye view grid over the ±153.6\pm153.6 m detection range. Each cell holds the log-count of returns in the cell and in its 3×33 \times3 and 5×55 \times5 neighbourhoods, mean radar cross-section and mean speed, giving a five-channel field EE. The field has no parameters and is recomputed per sample.

Per-query descriptor. The decoder carries each query as a ray: polar centre (θ,d)(\theta,d), height, log-size, heading and velocity. We convert each query to its metric centre cqc_q and mean predicted box dimension sqs_q in the same way the detector’s own consumers do, read the field there by bilinear sampling, eq=E(cq)e_q = E(c_q), and form, with ρq=∥cq,xy∥2\rho_q = \lVert c_{q,xy}\rVert_2 and ρ~q=clip⁡(ρq/153.6,0,2)\widetilde{\rho}_q = \operatorname{clip}(\rho_q/153.6,0,2),

dq=[eq,g(q)],g(q)=[ρ~q,(ρq+1)−1,sq/(ρq+1),log⁡(1+sq)],hq=MLP⁡(dq),(1)d_q = [e_q, g(q)], \qquad g(q) = [\widetilde{\rho}_q,(\rho_q+1)^{-1},s_q/(\rho_q+1),\log(1+s_q)], \qquad h_q = \operatorname{MLP}(d_q), \tag*{(1)}

where the MLP has two ReLU layers of 64 units.

Three branches. Three linear heads read hqh_q:

αq=0.6+1.9σ(wαThq+bα),πq=softmax⁡(Wπhq+bπ)∈Δ7,zq=wzThq+bz.(2)\alpha_q = 0.6 + 1.9\sigma(w_{\alpha}^{\mathsf{T}}h_q+b_\alpha), \qquad\pi_q = \operatorname{softmax}(W_{\pi}h_q+b_{\pi}) \in\Delta^7, \qquad z_q = w_z^{\mathsf{T}}h_q+b_z. \tag*{(2)}

The spatial branch scales the radial extent of the layer’s ray sampling by αq\alpha_q (cf. Zhu et al., [52]): the sampling points’ radial offsets, set per layer by RaCFormer, are multiplied by αq\alpha_q. The temporal branch multiplies the detector’s aggregation weight for frame tt by 8πq,t8\pi_{q,t}, whose mean over the eight frames is one, in the image and BEV samplers, without renormalising. It reads current-frame evidence and geometry only, so it learns a policy over frame offsets rather than measuring the evidence in each historical frame. The density branch is trained with

L=Ldet+0.1 BCE⁡(zq,σ(4(nq−0.5))),nq=sg⁡[E0(cq)],(3)\mathcal{L} = \mathcal{L}_{\mathrm{det}} + 0.1\,\operatorname{BCE}(z_q,\sigma(4(n_q-0.5))), \qquad n_q = \operatorname{sg}[E_0(c_q)], \tag*{(3)}

where E0E_0 is the log-count channel and sg stops the gradient; the loss covers the queries of the final decoder iteration. The target is derived from the input counts, so this is an auxiliary regulariser, not independent supervision; zqz_q feeds no prediction and acts on the detector only through the shared trunk. With zero-initialised head weights and bα=logit⁡(0.4/1.9)b_\alpha= \operatorname{logit}(0.4/1.9), training starts with αq=1\alpha_q = 1, uniform frame weights and zq=0z_q = 0, from a model that is bit-identical to the baseline.

A control module. An early version of EVISCOPE read the field at the raw polar values (θ,d)(\theta,d) instead of the decoded centre. Because these values address one fixed sliver of the grid for every query, the field samples do not represent the radar returns near each decoded centre; the four geometric terms remain query-dependent. The version trained without complaint and scored 26.11 mAP / 39.34 NDS. We keep it as the evidence-blind control of § 6.2: it has the same branches and the same per-query temporal weights, but no correctly aligned query-local evidence. It is an implementation-derived control evaluated at one seed, not a replicated geometry-only ablation. A separate geometry-only row is reported under the shortened one-epoch schedule (Appendix § 18); the full-schedule, three-seed geometry-only control is in Appendix § 30.

Cost. One module serves all six shared decoder iterations. It holds 5,450 of the model’s 64.0M parameters, 0.009%. Median latency is 965 ms against 997 ms for the baseline build and peak memory is 3.811 GB for both (Appendix § 20).

Experiments

Setup

We use MAN TruckScenes (Fent et al., [7]) under the official detection_cvpr_2024 protocol: 12 classes, centre-distance matching at {0.5,1,2,4}\{0.5,1,2,4\} m, cars scored to 150 m, following the protocol inherited from nuScenes (Caesar et al., [2]). We report the validation split. All models use camera and radar only, with no LiDAR or inference or as depth supervision. The detector is RaCFormer (Chu et al., [5]) with a ResNet-50 backbone (He et al., [10]) at 256×704256 \times704 input; every row we train uses class-balanced resampling (Zhu et al., [51]), eight temporal frames and four epochs on one node of eight A100 GPUs at an effective batch of eight. We call this configuration without the module the baseline, and the same configuration with the module EVISCOPE. Where we report several runs, the runs differ only in the random seed, which sets the initialisation and the data order; the samplers are not deterministic, so two runs at the same seed would also differ. Table 28 lists the settings of every comparison.

comparisonschedulerunswhat differs
Seven matched pairs (Table 7)4 ep7 + 7module on / off, same seed
Eight epochs (Table 8, Table 26)8 ep2 + 2module on / off, same seed
Geometry-only (Table 26)4 ep3field channels zeroed in dqd_q
Shuffled evidence (Table 26)4 ep2field channels permuted across queries of each sample
Density head (Table 11, Table 26)4 ep3density branch removed
Higher resolution (Table 1)8 ep1256×864256 \times864; no paired baseline
Component ablation (Table 11)4 ep1 eachcumulative branches; fixed radius, no module
Single-epoch suite (Table 12)1 ep, multi-frame from step 11 eachbranches, density head, evidence input
Recipe (Table 2)4 ep (12 ep without resampling)1 eachresampling, frame count
V2-99 pair (Table 9)4 ep1 + 1backbone
Substitutions (Table 6)trained weights4 EVISCOPE + 2 controlsone modulation replaced at inference
Held-out recall (Table 4)4 epseeds 5, 6module on / off; subset fixed from seeds 0–4
Probe (Table 5)trained weights4none (hooks only)
Scene conditions (Table 27)trained weights3 + 3official scene tags, ≥300\ge300 frames
Probe on the control (Figure 7)trained weights5 + 3none (hooks only)
Degradation (Table 24)trained weights1 + 1radar inputs dropped, same mask for both

Table 28. Summary of experimental settings for the comparisons in this paper (Table 26, Table 27, Figure 7). Runs counts completed trained models per arm used in the reported comparisons. Inference-time substitutions and probes re-evaluate trained weights without retraining.

rowradiuswindowevid. headevidence inputmAPNDS
baseline (no module)————14.2026.65
radius + window✓✓—full14.9527.58
radius only✓——full12.1325.77
window only—✓—full10.4624.33
full module✓✓✓full12.5422.67
full module✓✓✓geometry only12.8424.74

Table 12. Component ablation under a shortened single-epoch schedule. All rows share one recipe (CBGS, one epoch, multi-frame from the first step) and differ only in the module configuration, so the column is internally comparable; it is not comparable to Table 1, which uses the full four-epoch schedule. radius and window are the two modulation paths; evidence input denotes whether the descriptor sees the evidence field or only its geometric terms.

The training recipe

Our ResNet-50 baseline reaches 25.9 mAP over seven runs against 24.1 for the highest published mAP in our comparison, with a standard deviation of 0.45 mAP across our runs. Table 1 places it among the published entries on TruckScenes val, as reported by Bang et al. [1] ([48, 16, 20, 17, 28, 46]). In NDS the four-epoch baseline is at 38.1 against 38.4; repeated runs of the published models would be needed to assess this difference. The error terms are not uniformly better: translation and attribute errors improve over the previous best, scale is close, orientation and velocity are worse.

Table 1: TruckScenes val. L: LiDAR, C: camera, R: radar; bold marks the best camera–radar value in a column, the LiDAR row being a reference. Four-epoch rows are seven-run means (Table 7), eight-epoch rows one run each; the margin over the published entries is present in the module-free baseline rows. paired Δ is the within-seed difference, not the difference of the rounded means: the seven-pair mean (±0.36 NDS, ±0.22 mAP; § 5.3) and the single eight-epoch pair. A V2-99 pair is in Appendix § 14. ‡Error terms not retrieved (Appendix § 13). §One run at the published ResNet-50 input size, without a paired baseline (Appendix § 13).

MethodIn.BackboneImg sizeEp.NDS↑mAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓
CenterPoint-VLVoxel——35.322.60.4610.4050.4683.0280.261
Far3DCV2-99640 × 960—21.410.70.8830.5070.6711.3520.338
RCTransC+RR50256 × 864—22.912.61.0250.5070.6080.8940.331
CRT-FusionC+RR50256 × 864—28.816.90.8330.5120.4440.8520.322
BEVFusionC+RR50256 × 864—30.418.20.9410.4420.3890.8920.208
SpaRCC+RV2-99640 × 960—35.422.50.7980.4490.4760.6130.248
Horizon3DC+RR50256 × 8644837.423.60.8760.4100.3310.6250.198
Horizon3DC+RV2-99640 × 9604838.424.10.8330.4040.3550.5610.208
Ours (baseline)C+RR50256 × 704438.125.90.6550.4180.5370.7470.123
Ours (EVISCOPE)C+RR50256 × 704438.425.60.6500.4180.5190.7220.126
paired Δ4+0.31−0.27
Ours (baseline)‡C+RR50256 × 704839.226.4—————
Ours (EVISCOPE)C+RR50256 × 704839.926.70.6540.4080.4800.6710.129
paired Δ8+0.67+0.26
Ours (EVISCOPE)§C+RR50256 × 864839.928.20.5990.4120.5680.7190.127

Table 1.

By range. Table 3 restricts the protocol to disjoint bands. Both seed-0 models exceed the published ResNet-50 entry in the two nearest bands. At 100–150 m, the module arm reaches 9.4 mAP / 19.2 NDS against 8.9 / 18.4, while the baseline reaches 8.6 / 18.1; the published entry leads at 50–100 m. These are single runs.

Table 3: mAP / NDS by disjoint range band, official protocol restricted to each band. Published rows are from Bang et al. [1], Table 9; our rows are the seed-0 models. Bold marks the best value in each column.

0–25 m25–50 m50–100 m100–150 m
mAPNDSmAPNDSmAPNDSmAPNDS
RCTrans19.928.412.524.17.919.96.215.6
CRT-Fusion26.234.416.527.311.026.37.011.4
BEVFusion28.835.716.825.211.720.17.413.9
Horizon3D (R50)36.543.023.337.817.032.78.918.4
Ours (baseline)41.244.525.339.015.729.28.618.1
Ours (EVISCOPE)43.047.025.339.416.830.59.419.2

Table 3.

Training budget. The previous best trains for 48 epochs. Doubling ours to eight, with every other setting fixed, lifts both arms above it on both headline metrics: 26.4 mAP / 39.2 NDS without the module and 26.7 / 39.9 with it. Epoch counts are not a compute-normalised comparison: resampling, iterations per epoch, batch sizes and input sizes differ, and we do not report matched total iterations or GPU-hours for the published entries. Against the four-epoch mean the module arm gains 1.08 mAP and 1.47 NDS, with orientation and velocity errors falling to 0.480 and 0.671 (Appendix § 13). These are single runs, and the NDS margin (0.8–1.5) is the size of the spread between four-epoch runs (§ 6.3); the remaining gap to the previous best is in orientation and velocity.

Table 2 examines class-balanced resampling and temporal window length. The no-CBGS run is 2.75 mAP below the baseline mean, but also changes the schedule and sample exposure, so it does not isolate the effect of CBGS. The four- and two-frame runs are lower by 1.83 and 3.49 mAP, respectively, with velocity error rising from 0.754 to 0.957. These single-run results are consistent with benefits from longer temporal context, as also observed in camera-only detectors [13, 38, 44]. They do not establish whether the benefit comes from radar returns, image context or motion estimation, nor do they quantify each recipe choice’s contribution across seeds.

Table 2: Recipe ablation on the baseline (no module), ResNet-50. The first row is the seven-run mean of Table 7; the others are single runs. †Without class-balanced resampling an epoch has about a quarter the iterations, so this run was given 12 epochs with the same fraction of training before the multi-frame phase; it still sees about 30% fewer samples and differs from the first row in more than one setting.

configurationepochsmAP↑NDS↑mAVE↓runs
CBGS, 8 frames (baseline)425.8538.130.7547
no CBGS, 8 frames†1223.1035.690.7131
CBGS, 4 frames424.0236.240.8441
CBGS, 2 frames422.3633.910.9571

Table 2.

Seven paired runs do not separate the module from zero

Figure 1b shows the baseline and EVISCOPE trained seven times each from matched seeds. Two things are visible at once. The seed moves both models: their NDS across seeds correlates at r=0.79r = 0.79, and the spread across seeds (1.5 NDS for the baseline, 3.7 for EVISCOPE) is larger than any within-pair difference after the first two seeds. And the within-pair difference has no consistent sign: the seven paired NDS differences, printed above each pair in Figure 1b and tabulated with the per-seed scores in Appendix § 12, average +0.31 with a standard error of 0.36 (t=0.87t = 0.87 on six degrees of freedom), and the paired mAP difference is −0.27 ± 0.22. The NDS point estimate is smaller than its standard error.

We report within-seed differences because the seed is a shared source of variation ([40, 6]). With the same seven seeds in both arms, the mean paired difference equals the difference of means; pairing matters for the uncertainty estimate. These estimates do not establish an improvement or equivalence between the models. Recall, a count over 86,252 objects that does not depend on score calibration, is a quieter statistic. Its paired difference over the seven seeds is +0.12 ± 0.14 points (t=0.86t = 0.86).

The first two pairs alone average about +1.5 NDS, substantially above the seven-pair estimate and illustrating the sensitivity to which runs are reported.

The module starts at identity and can only act once training has moved it off that point. The reported differences are −3.98 NDS under the shortened one-epoch schedule, +0.31 ± 0.36 at four epochs, and +0.67 NDS / +0.26 mAP for one pair at eight. Because the one-epoch schedule changes multi-frame activation and the budgets have different numbers of runs, these values do not establish a training-budget trend. We had fixed beforehand that one pair cannot settle this and that only a difference above +1.0 would justify more pairs (Appendix § 11); +0.67 is recorded as an intermediate result. During review we completed one further pair at a reviewer’s request (Appendix § 30).

The far-range recall gain does not survive a held-out test

On the seed-0 pair, the module’s recall gain was not spread evenly. It was 0.6–1.4 points in every hit bin over all ranges, and 4.3 and 6.2 points for objects at 100–150 m with 3–5 and 6–15 returns (Appendix § 16). This was an appealing pattern: distant objects with a few returns are where a wider search and a longer temporal window should help most. When we had five pairs, four of them showed a positive difference on this subset, and the mean was +1.68 points.

Because we had chosen the subset after looking at the data ([8, 41]), we wrote down a confirmation test before computing recall for the two remaining pairs (Nosek et al., 2018). Here “held out” refers to training seeds; all pairs use the same validation objects. The internally recorded protocol is reproduced in Appendix § 11. The subset was fixed as the two cells merged by count, the matching threshold at 2 m, and the criterion as both held-out differences being positive. Table 4 gives the result. The held-out pairs give −0.67-0.67 and −0.59-0.59. The seven-pair mean is +1.02±0.99+1.02 \pm0.99. The gain does not satisfy the pre-specified replication criterion. This limits the interpretation of the exploratory subset, without establishing a zero population effect.

Table 4: Paired recall difference (EVISCOPE minus baseline, points) on objects at 100–150 m with 3–15 returns, the subset where the seed-0 pair gained most (Appendix § 16). Seeds 0–4 were examined before the test was written; the internally pre-specified criterion was that both held-out differences be positive.

s0s1s2s3s4s5 (held out)s6 (held out)
Δ\Deltarecall+4.96+0.51−2.64+3.27+2.32−0.67−0.59

Table 4.

Analysis

What the module learns, and what an ablation of it measures, are separate questions. We ask whether it learned what it was meant to learn, what the standard attribution test says about it and whether that test can be trusted, and how much a single run on this benchmark can resolve.

The module learns the intended allocation

Table 5 reports the allocation the trained model makes. The sampling radius decreases across the local-density bins and increases across the range bins. The temporal window is similar in the first two range bands, rises through 100–150 m, and falls in the final band. The rank correlation of radius with local evidence is ρ=−0.91\rho= -0.91 and with range +0.51+0.51; these marginal correlations do not separate their contributions. The same probe on three further independently trained runs gives −0.90-0.90, −0.92-0.92 and −0.91-0.91. The ordering is in place within the first epoch (Appendix § 24). These four probes confirm an evidence-associated allocation, while the paired comparisons do not establish an end-to-end benefit. The observed radii fall in [0.60,0.84][0.60, 0.84], below the initial value of 1: the module shrinks the sampling extent overall and adds an evidence-dependent spread on top of that.

Inference-time dependence does not establish incremental value

A common way to attribute a model’s behaviour to one component is to disable that component at inference and measure the drop (Zhou et al., 2018; Morcos et al., 2018; Voita et al., 2019; Michel et al., 2019). We did this for both of the module’s outputs. We take a trained EviSCOPE model and, at inference only, replace the radial scale by its per-frame mean, by a shuffled copy of itself, or by a constant, or replace the temporal weights by uniform ones, which is what the baseline uses. Everything else is untouched, and we verify that each intervention fires by counting calls and recording the standard deviation of the radial scale before and after (Appendix § 21).

The two channels look very different under this test. Table 6 shows the result. Replacing the radial scale by its per-frame mean changes NDS by −0.03 ± 0.18 and recall by +0.06 ± 0.07 points across four models (mean ± SD). Other radial substitutions also have much smaller observed effects than temporal substitution (Appendix § 26). Making the temporal weights uniform reduces performance in all four models, by 11.0 ± 2.0 NDS (mean ± SD), 11.4 mAP and 8.2 recall points on average, and by 10.45 NDS on the V2-99 model, whose end-to-end comparison runs the other way. This contrast measures dependence on the trained outputs, not the benefit of introducing either branch.

modelsubstitutionΔ\Delta mAPΔ\Delta NDSΔ\Delta recall
EviSCOPE, 4 runsradial scale →\to per-frame mean+0.00±0.11+0.00 \pm0.11−0.03±0.18-0.03 \pm0.18+0.06±0.07+0.06 \pm0.07
temporal weights →\to uniform−11.4±3.2-11.4 \pm3.2−11.0±2.0-11.0 \pm2.0−8.2±2.6-8.2 \pm2.6
EviSCOPE, V2-99 evidence-blind controltemporal weights →\to uniform−13.23-13.23−10.45-10.45−9.81-9.81
temporal weights →\to uniform−8.71-8.71−11.75-11.75—

Table 6. Inference-time substitution: each row replaces one modulation on a trained model and re-evaluates the same weights, with no retraining. The first block is mean ±\pm standard deviation over four ResNet-50 EviSCOPE models; the remaining models are one run each. recall is the fraction of evaluated objects matched within 2 m. The further substitutions and the control model of § 6.2 are in Appendix § 26.

Dependence also occurs without query-local evidence. Two further models change the reading. The first is the module trained without the density branch, which removes the auxiliary supervision on the trunk; making its temporal weights uniform still costs 7.34 NDS (Appendix § 26), so this dependence can arise without density supervision. The second is the evidence-blind control of § 4.2, whose descriptor lacks correctly aligned local radar information and whose spatial branch, we find, has collapsed to a constant (its standard deviation across queries is 0.000). Making its temporal weights uniform costs 11.75 NDS. This control also depends strongly on its temporal weights; its single run does not establish equivalence to the correctly conditioned module.

What the test measures. The baseline is trained with uniform temporal weights from the start and reaches 38.13 NDS over seven runs. A model trained with learned per-query weights and then evaluated with uniform ones reaches 25–32. This is consistent with co-adaptation to the learned weights ([11]); substitution changes the inputs to downstream computation without allowing it to adapt. It does not isolate the value of the evidence signal. Related concerns motivate retraining in feature attribution ([12]), pruning ([29]) and architecture selection ([43]). Table 7 estimates the complete module’s incremental effect. Isolating the evidence input requires a deliberately constructed geometry-only conditioner trained under the same full schedule and matched seeds; Appendix § 30 reports it at three seeds.

Table 7: The ResNet-50 pair at seven seeds, in the configuration of § 5.1; Δ\Delta is EVISCOPE minus baseline within a seed. Seeds 5 and 6 were trained after the held-out test of § 5.4 was written down.

s0s1s2s3s4s5s6mean
mAPmAP
baseline25.8426.4826.1725.0025.8625.8225.7625.85
EVISCOPE26.4726.5125.7724.4225.7924.5625.5225.58
Δ\Delta+0.63+0.03−0.40−0.58−0.07−1.26−0.24−0.27
NDSNDS
baseline38.4838.8537.3937.3438.6037.5538.6938.13
EVISCOPE40.1940.2437.7337.4438.1636.5438.8038.44
Δ\Delta+1.71+1.39+0.34+0.10−0.44−1.01+0.11+0.31

Table 7.

Training-time controls. The radial channel could also act during training and be redundant afterwards. Models trained from scratch with every radial scale fixed at 0.6, the value the trained module most often emits, reach 38.02, 37.63 and 37.37 NDS at three seeds, 0.46, 1.22 and 0.02 below their paired baselines. These controls contain no modulation module, so they do not isolate radial adaptation while retaining learned temporal weights; isolating these branches remains an open comparison. The seed-0 component ablation (Appendix § 17) is also insufficient to establish each branch’s contribution across seeds.

A single run cannot resolve one point of NDS

Across seeds in the configuration of § 5.1, our seven baseline runs span 1.51 NDS (standard deviation 0.67) and our seven EviSCOPE runs span 3.70 (1.39). The two best published entries differ by 1.0 NDS, and every published entry is a single run, which cannot resolve a difference of that size.

Radar degradation. Re-scoring the seed-0 pair on identical corrupted inputs tests sensitivity to radar corruption (Appendix § 27). Thinning the returns, uniformly or at long range, leaves the within-pair difference between +0.55 and +0.70 mAP at every rate.

The score is concentrated by scene: highway frames reach 38.1 NDS against 22–25 for city and terminal, so a val mean mixes regimes that differ by more than any effect here (Appendix § 23).

The observed standard deviation is about twice as large for the module arm; the seed-2 EviSCOPE run is two NDS below seed 0 after the first epoch and remains lower at the reported checkpoints (Appendix § 24).

With a paired standard deviation of 0.96 NDS and an effect of +0.31, about 40 pairs would bring the mean to twice its standard error and about 80 would give 80% power ([3]), roughly 1,200 GPU-days (Appendix § 15); we report the seven pairs as they stand.

Conclusion

We introduced EviSCOPE to test radar evidence as an additional signal for query sampling in camera–radar detection. On nuScenes, the module-free recipe exceeds the published mAP values in our comparison, while seven matched pairs leave the module’s incremental benefit unresolved: +0.31±0.36+0.31 \pm0.36 NDS and −0.27±0.22-0.27 \pm0.22 mAP (mean ±\pm SE). Against a geometry-only conditioner of the same size, EVISCOPE is higher by +0.87±0.21+0.87 \pm0.21 mAP and +1.25±0.72+1.25 \pm0.72 NDS over three matched seeds. An exploratory subset gain fails an internally pre-specified test on two further seeds, and large temporal substitution penalties also occur without correctly aligned query-local evidence. These findings support evaluating learned allocation, inference-time dependence and paired detection gains separately. Evaluation beyond this benchmark remains necessary to assess the broader hypothesis.

Ethics Statement

This work uses the public MAN TruckScenes dataset under its research licence; it involves no human subjects and no personal data beyond what the dataset releases with faces and licence plates anonymised. The application domain, perception for autonomous trucks, is safety-relevant. We report sensitivity to simulated radar-sector failure (Appendix § 27) and note that the evidence descriptor does not explicitly encode sensor coverage, so that this limitation can be considered before deployment.

Reproducibility Statement

§ 5.1 describes the data splits and evaluation settings. Appendix § 21 documents data preparation, hyperparameters, training procedures, and implementation differences between comparison models. Diagnostic definitions and thresholds appear in § 4.1, § 6.1, and § 6.2, with component ablations and additional diagnostics in Appendices § 17, § 18, § 26, and § 27. Validation-set results are explicitly identified, and exploratory and held-out analyses are distinguished in § 5.4 and Appendix § 11. Training and evaluation code is included in the supplementary material and will be publicly released.

AI Use Statement

In this work, we used generative AI tools only to edit author-written text for grammar, style, and readability. The authors reviewed all AI-assisted revisions and take full responsibility for the final manuscript, including its claims, citations, and experimental reporting.

Internally pre-specified held-out test

The far-range low-evidence subset of § 5.4 was found on the seed-0 pair and examined on seeds 1–4. Before recall was computed for the seed-5 and seed-6 pairs, we fixed the following, reproduced here from our laboratory log with only the wording translated. The held-out units are training seeds evaluated on the same validation split; the description here documents an internal protocol.

Origin. The subset was identified on the seed-0 recall table (Table 10) and has been examined on seeds 1–4; this is exploratory. Recall has never been computed for the seed-5 and seed-6 pairs; they form the held-out set.

range0 hits1–23–56–15>15
0–25 m+1.5−0.6+0.7+0.3+1.7
25–50 m−0.6+1.5−1.1+0.7+0.3
50–100 m+1.1+1.4+1.2+2.1+1.1
100–150 m+0.8+0.5+4.3+6.2—
all ranges+0.6+0.9+0.8+1.4+0.9

Table 10. EVISCOPE minus baseline recall, in points, by range band and radar-hit bin, seed-0 pair only, same protocol as Table 20. The >15 cell beyond 100 m holds 23 objects and is omitted. The two bold cells did not replicate on held-out pairs.

Hypothesis. EVISCOPE improves recall over the same-seed baseline on objects at 100–150 m with 3–15 radar returns, taking the 3–5 and 6–15 cells merged by object count as one recall rate.

Criterion. If the merged recall difference is positive on both held-out pairs, the result is recorded as replicated; if either is ≤0\le0, as not replicated. The seven-pair estimate (mean ±\pm SE) is reported as well, but the conclusion rests on the two held-out pairs only. Overall recall and other cells are not criteria for this test.

Not permitted. Changing the cells, the range boundaries, the matching threshold (fixed at 2 m), or dropping any seed. The result is reported either way.

The held-out differences were −0.67 and −0.59 (Table 4). The test was recorded as not replicated.

Training-budget hypothesis. Before the eight-epoch pair of § 5.3 was trained we recorded the following.

Motivation. The module is initialised at identity and must leave that initialisation before it can act. The paired difference is monotone in budget: −1.66 mAP / −3.98 NDS at one epoch, +0.31 ± 0.36 NDS over seven pairs at four. The published best trains for 48 epochs; we have four.

Hypothesis. Raising the budget from four to eight epochs increases the paired Δ\DeltaNDS above +0.31.

Design. The two eight-epoch runs are the existing four-epoch configurations of both arms with the epoch count changed from 4 to 8 and nothing else; same hardware, same batch size per device, same seed (0).

Criterion. The eight-epoch paired Δ\DeltaNDS decides. One pair cannot settle the question (run-to-run spread 1.5–3.7 NDS). If Δ\DeltaNDS > +1.0, two further pairs are warranted; if ≤0\le0, the budget hypothesis is recorded as not supported and no further pairs are trained.

Not permitted. Changing the epoch count after seeing the result and reporting selectively; unequal budgets between arms.

The pair gave +0.67 NDS / +0.26 mAP, between the two thresholds; it is recorded as an intermediate result and no claim about the module rests on it. The motivation above records our hypothesis at the time. The one-epoch experiment also changes the multi-frame schedule, and only one pair was run at eight epochs, so these results do not establish a budget-dependent effect. The additional completed eight-epoch pair of Appendix T was added during review at a reviewer’s request, not because this criterion called for it.

The seven seed-matched pairs

Training budget

Table 8 gives the eight-epoch runs of Table 1 in full, next to the seven-run four-epoch means. Both arms are one run. The baseline was evaluated on four GPUs and the module run on eight, on different machines; evaluating one trained model under both set-ups differs by about 0.12 NDS, with the module’s machine reading lower. Error terms for the eight-epoch baseline were not retrieved at the time of writing. One further run trains EVISCOPE for eight epochs at 256 × 864, the input size of the published ResNet-50 entry, and reaches 28.22 mAP / 39.86 NDS (mATE 0.599, mASE 0.412, mAOE 0.568, mAVE 0.719, mAAE 0.127). It has no paired baseline and says nothing about the module. We had predicted that the larger input would bring orientation error below 0.47; it did not, and we did not pursue resolution further. An earlier run at this size changed hardware and per-GPU batch mid-training and is discarded (Appendix § 21).

Table 8: Four- versus eight-epoch training, ResNet-50 at 256×704256 \times704, with the other training settings fixed. Four-epoch rows are seven-run means (Table 7); eight-epoch rows are single runs evaluated on different GPU configurations, as described above.

modelepochsmAP↑mAP^{\uparrow}NDS↑NDS^{\uparrow}mATE↓mATE^{\downarrow}mASE↓mASE^{\downarrow}mAOE↓mAOE^{\downarrow}mAVE↓mAVE^{\downarrow}mAAE↓mAAE^{\downarrow}
baseline425.8538.130.6550.4180.5370.7470.123
baseline826.4039.24—————
EVISCOPE425.5838.440.6500.4180.5190.7220.126
EVISCOPE826.6639.910.6540.4080.4800.6710.129

Table 8.

The V2-99 pair

We trained one pair on the V2-99 backbone. The baseline of that pair is our strongest same-backbone comparison: at 256×704256 \times704 and four epochs it exceeds the previous best on both headline metrics (Table 1), against a model of the same backbone trained at 640×960640 \times960 for 48 epochs. Table 9 gives both runs. The within-pair difference, −0.1-0.1 mAP and −1.1-1.1 NDS, is one pair and falls inside the single-run spread of § 6.3; the seven-pair estimate of § 5.3 is the measurement we read for the module.

Table 9: The V2-99 pair at 256×704256 \times704 and four epochs, seed 0, single runs. Every other setting is that of § 5.1.

modelmAP↑mAP^{\uparrow}NDS↑NDS^{\uparrow}mATE↓mATE^{\downarrow}mASE↓mASE^{\downarrow}mAOE↓mAOE^{\downarrow}mAVE↓mAVE^{\downarrow}mAAE↓mAAE^{\downarrow}
baseline26.739.00.6430.4100.5060.7520.124
EVISCOPE26.637.90.6510.4100.5300.8140.135

Table 9.

What it would take to resolve the effect

With a paired standard deviation of 0.96 NDS and an effect of +0.31, about 40 pairs would bring the mean to twice its standard error and about 80 would give 80% power ([3]), roughly 1,200 GPU-days. That budget is out of proportion to the effect, and we report the seven pairs as they stand.

Exploratory recall gain on the seed-0 pair

Table 10 is the table on which the far-range subset was found. It is the difference between Table 21 and Table 20 for the seed-0 pair, and it is a single pair; § 5.4 and Appendix § 11 give the outcome of testing its two largest cells on further pairs.

0 hits1–23–56–15>15
0–25 m83.588.794.196.491.9
25–50 m69.584.788.088.992.7
50–100 m70.083.082.080.284.9
100–150 m73.081.478.577.065.2
0–25 m3,8854,0423,4554,4511,291
25–50 m9,9506,5123,4122,5421,341
50–100 m14,2598,6883,4352,682741
100–150 m8,7584,2431,64389923

Table 21. EVISCOPE recall (%) by range and by radar returns inside the box (top), and the number of val objects in each cell (bottom); same protocol as Table 20.

Table 10: EVISCOPE minus baseline recall, in points, by range band and radar-hit bin, seed-0 pair only, same protocol as Table 20. The >15 cell beyond 100 m holds 23 objects and is omitted. The two bold cells did not replicate on held-out pairs.

Component ablation at four epochs

Table 11 splits the module at the full schedule, at seed 0. Every row is a single run. The differences among the first three rows are smaller than the run-to-run spread of § 6.3, and we do not interpret them; we keep the table because dropping single runs after seeing them would be selective reporting. The fixed-radius rows are the training-time control discussed in § 6.2; a third run at seed 2 scored 25.92 mAP / 37.37 NDS.

Single-epoch ablation

The single-epoch suite we ran before the four-epoch ablation of Appendix § 17. All rows use one recipe (CBGS, one epoch, multi-frame from the first step); the column is internally comparable and not comparable to the four-epoch tables. The full module’s difference from the baseline and the increment from adding the density head reverse sign between these schedules, while radius plus window improves both reported metrics in both. These single-run results do not support extrapolating component ablations from the shortened schedule.

Table 12: Component ablation under a shortened single-epoch schedule. All rows share one recipe (CBGS, one epoch, multi-frame from the first step) and differ only in the module configuration, so the column is internally comparable; it is not comparable to Table 1, which uses the full four-epoch schedule. radius and window are the two modulation paths; evidence input denotes whether the descriptor sees the evidence field or only its geometric terms.

Table 13: Per-class AP change from adding the module (EVISCOPE minus baseline, same recipe), seed-0 pair only. Per-class differences of this size are within the run-to-run spread of Section 5.3 and we do not read them. Absolute per-class values are in Appendix I.

cartruckbustrail.o.veh.ped.moto.bikeconebarr.anim.sign
Δ\DeltaAP+0.7+1.9+1.9+1.3−4.1+2.1−0.3+1.2+2.5−1.30.0+1.7

Table 13.

Table 14: Per-class AP (%) on TruckScenes val for the rows we train. Rows share the recipe of Table 1; seed 0 unless stated.

cartruckbustrail.o.veh.ped.moto.bikeconebarr.anim.sign
baseline (R50)57.837.50.149.011.010.446.612.543.32.90.039.0
+ radius & window (R50)58.038.00.549.411.310.646.811.343.59.50.038.7
EVISCOPE (R50)58.539.42.050.36.912.546.313.745.81.60.040.7
baseline (R50, seed 1)58.239.25.351.39.39.448.312.041.04.50.039.4
EVISCOPE (R50, seed 1)58.738.30.851.512.07.947.314.840.96.90.039.1
baseline (R50, seed 2)58.037.61.551.08.110.243.710.040.515.10.038.5
EVISCOPE (R50, seed 2)58.237.41.749.74.211.850.912.740.41.70.040.6
baseline (V2-99)58.238.60.751.110.212.849.313.144.81.10.040.9
EVISCOPE (V2-99)59.138.31.851.29.612.646.111.842.35.20.041.4

Table 14.

Absolute per-class AP

Efficiency

Measured on one A100 at batch one with the full eight-frame pipeline, after warm-up, over 50 timed iterations; the two builds differ only in whether the module is attached. Absolute latency includes data preparation on the host and is not comparable to throughput figures reported elsewhere under other protocols; the difference between the two rows is the module’s cost. Analytically the module costs 5,312 multiply-adds per query and decoder iteration, about 29 M per sample for 900 queries and six iterations, plus one scatter of the radar returns into the 128×128128 \times128 field.

parametersof which modulemedian latency (ms)peak memory (GB)
baseline63,998,04009973.811
EVISCOPE64,003,4905,4509653.811

Table 15.

Implementation details

Detector. RaCFormer [5] with a ResNet-50 backbone initialised from nuImages weights, FPN neck, input 256×704256 \times704, 150 rays ×\times 6 range clusters = 900 queries, six shared decoder iterations with per-layer sampling extents {0.08,0.07,0.06,0.05,0.04,0.03}\{0.08, 0.07, 0.06, 0.05, 0.04, 0.03\} of the scene, eight frames (one current, seven past), radar aggregated over five sweeps per frame. Training uses class-balanced resampling [51], an effective batch of 8, AdamW at 2×10−42 \times10^{-4} with cosine decay, four epochs, with multi-frame input enabled from the second epoch. The single-epoch rows of Table 12 enable multi-frame input from the first step and stop after one epoch.

Module. Evidence grid 128×128128 \times128 over ±153.6\pm153.6 m; count channels at dilations of 0, 1 and 2 cells, plus mean cross-section and mean speed; descriptor MLP with two hidden layers of 64; radius scale αq∈[0.6,2.5]\alpha_q \in[0.6, 2.5] from a sigmoid; frame weights from a softmax over 8 frames scaled by 8; density head trained with binary cross-entropy against σ(4(nq−τ))\sigma(4(n_q-\tau)) ( Equation (3)), where nqn_q is the detached log-count at the query’s centre and τ=0.5\tau= 0.5 is a fixed constant. The auxiliary loss is weighted 0.1 against the detection loss and covers the final decoder iteration only, because the shared layer overwrites the earlier iterations’ logits. Parameters: 9⋅64+649 \cdot64+64 and 64⋅64+6464 \cdot64+64 in the trunk, 65, 520 and 65 in the three heads, 5,450 in total.

The implementation computes zqz_q at inference, but detection does not consume it. Removing this unused head computation leaves the detections unchanged.

componentsource
Camera branch: ResNet-50, FPN, LSS depth and BEV view transformRaCFormer
Radar branch: pillar encoder, radar BEV, radar depthRaCFormer
Polar query initialisation (150 rays × 6 clusters), query denoisingRaCFormer
Shared decoder layer (6 iterations): ray sampling in image, LSS-BEV and radar BEV; per-frame aggregation weightsRaCFormer
Heads, assigner and lossesRaCFormer
Four cameras, 150 m range, 256 × 256 BEV grid, TruckScenes class mappingadaptation
Recipe: class-balanced resampling, four (eight) epochs, multi-frame input from epoch twoadaptation
Evidence field EE (5 × 128 × 128, no parameters), Equation (1)EViSCOPE
Descriptor dqd_q and 64-64 MLP, shared by the six iterationsEViSCOPE
Radial scale αq\alpha_q on the ray-sampling extentEViSCOPE
Temporal weights 8πq,t8\pi_{q,t} on the per-frame aggregation weightsEViSCOPE
Density head zqz_q (unused for detection; removable at inference) and auxiliary lossEViSCOPE

Table 15 (PDF p. 19). Inherited and added components. RaCFormer: unchanged from Chu et al. [5]. Adaptation: changes needed to run on TruckScenes, shared by baseline and EVISCOPE; not a method claim. EVISCOPE: the module of § 4.2, the only difference between the paired rows of Table 7.

Algorithm 1 One decoder iteration with EVISCOPE. Lines marked ▷ are the module; all others are RaCFormer.

Require: ray queries; image, LSS-BEV and radar-BEV features of 8 frames; field EE built once per sample from current-frame returns

  1. for all queries qq do

  2. (cq,sq)←(c_q, s_q) \leftarrow metric centre and extent of qq ▷ RaCFormer’s conversion

  3. ▷ hq←MLP⁡(dq)h_q \leftarrow\operatorname{MLP}(d_q); (αq,πq,zq)←heads⁡(hq)(\alpha_q, \pi_q, z_q) \leftarrow\operatorname{heads}(h_q) ▷ Eqs. (1)–(2)

  4. end for

  5. ▷ scale each query’s radial sampling extent by αq\alpha_q

  6. sample image, LSS-BEV and radar-BEV features along each ray

  7. ▷ multiply the per-frame aggregation weights of query qq by 8πq,t8\pi_{q,t}

  8. aggregate, refine the queries, predict boxes

  9. if training and final iteration then

  10. ▷ add 0.1BCE⁡(zq,σ(4(nq−0.5)))0.1\operatorname{BCE}(z_q,\sigma(4(n_q-0.5))) ▷ Eq. (3)

  11. end if

Activation recomputation. The samplers recompute their activations during the backward pass to save memory. A modulation reached through shared mutable state is re-read at recomputation time, by which point the shared decoder layer has overwritten it with the previous iteration’s value, so the backward pass differentiates a function other than the one the forward pass evaluated. We therefore pass the radial scale and the temporal weights to the samplers as explicit tensor arguments, which fixes their values for both passes.

Reading the field at the decoded centre. The decoder carries queries in polar form, [θ,d][\theta, d] normalised to [0,1][0, 1], and the metric centre is recovered by two conversions. The evidence-blind control of § 4.2 sampled the evidence field at the raw polar values instead, which addresses one fixed sliver of the grid for every query. Its descriptor therefore lacked correctly aligned query-local radar information, yet training converged and the model scored 26.11 mAP / 39.34 NDS at seed 0, between the baseline (38.48) and the corrected module (40.19) at the same seed. When we later hooked its radial head we found that it emits the same value for every query (standard deviation 0.000 across queries), despite retaining geometric inputs. This observation does not identify why the radial head became constant. The probe of Table 5 is the check that the corrected module reads what it is meant to read; we recommend it to anyone reimplementing a module of this kind, because an evidence-blind one trains without complaint.

local evidenceradiusrangeradiuswindow
[0, 0.35)0.7340–25 m0.6275.00
[0.35, 0.7)0.62925–50 m0.6374.94
[0.7, 1.1)0.61150–100 m0.6635.22
≥1.1\ge1.10.602100–150 m0.6925.45
> 150 m0.7995.12

Table 5. Forward-hook probe on the trained EviSCOPE model (seed 0), last decoder layer, 270,000 queries over 300 val samples. radius multiplies the layer’s sampling extent; window is exp⁡(H)\operatorname{exp}(H) of the temporal weights, in frames.

Single-node training. Every row is trained end-to-end on one node of eight A100 GPUs. We record one run we discarded: a seed-1 model whose first epoch was trained on Turing GPUs (FP32 matrix multiply) and whose remaining epochs were resumed on A100 (TF32) scored 37.92 NDS, against 40.24 for the same seed trained on A100 throughout. When we discarded it we read the gap as a numerics change; the seed-2 pair later showed that a seed alone moves both models by a similar amount (Table 7), so the discarded run may have been an unlucky draw rather than a corrupted one. We keep the rule regardless: no row in this paper mixes hardware or per-GPU batch size within a run. A second discarded run, at 256 × 864 input, changed both hardware and per-GPU batch size at the same resume point and scored 24.29 / 36.33; it is comparable to no row and is not reported.

Probe. Forward hooks on the last decoder iteration over 300 val samples (270,000 queries), reporting Spearman rank correlation between the emitted radius scale and, respectively, local return density at the query’s decoded centre and the query’s range.

Degradation. For each val sample the dropout mask is generated from a random seed derived from the sample token, so the two models in Table 24 receive identical inputs. random drops each return independently; far drops returns beyond 75 m at the stated rate; sector disables two of the six radar sensors.

degradationmAP EviSCOPEmAP baselineΔ\Deltarelative drop EviSCOPErelative drop baseline
none26.4825.83+0.65——
random drop 25%26.1625.46+0.70−1.2%−1.4%
random drop 50%24.9424.39+0.55−5.8%−5.6%
random drop 75%21.1120.42+0.69−20.3%−20.9%
far-range drop 50%26.2325.61+0.61−0.9%−0.8%
sector failure 33%21.7222.00−0.28−18.0%−14.8%

Table 24.

Dataset statistics

Table 16 extends the pooled row of Table 20 from the val split to the full trainval annotation set (774,688 boxes of the twelve official classes). An object is zero-return when no point of the current radar frame falls inside its box. The largest spread across weather or time-of-day tags in any band is 10 points; across range it is 57 points for cars, and across classes at 100–150 m it is 82 points. These marginal summaries do not isolate the effects of weather, class or scene composition.

Table 16: Fraction (%) of annotated objects with no radar return in the current frame, trainval split, against range: by class (top), weather tag (middle) and time-of-day tag (bottom). Cells with fewer than 200 objects are blank; the all classes row is the count-weighted mean.

0–25 m25–50 m50–100 m100–150 m
car15.732.051.872.6
truck3.98.111.123.4
bus0.27.712.215.8
trailer2.16.410.123.6
other vehicle0.37.97.624.9
pedestrian40.170.486.297.9
motorcycle–37.878.593.7
bicycle34.358.382.7–
traffic cone74.989.294.897.3
barrier70.978.571.9–
animal–70.180.2–
traffic sign56.372.381.487.2
all classes23.140.051.866.2
clear22.538.950.364.2
overcast24.543.555.369.0
rain22.036.349.566.0
snow26.044.748.561.8
fog19.938.349.364.1
morning21.736.947.761.8
noon22.741.454.070.0
evening31.743.055.767.7

Table 16.

Performance by scene tag

TruckScenes attaches scene-level tags for area, weather, lighting and time of day. Table 17 evaluates the EviScope model (seed 0) on the val frames carrying each tag, under the official protocol restricted to that subset. We show the tags with at least 100 val frames; val has no frame tagged night, and fog, snow and tunnel carry 40 frames each (two scenes), too few to report. Scene type dominates: highway frames score 38.1 NDS, city and terminal frames 22–25. The weather tags span 28–41 NDS, but tags are not independent of scene type (the rain frames come from a different mix of routes), so the table describes where the model is strong rather than isolating a cause.

Table 17: EVISCOPE (R50, seed 0) on val subsets defined by the official scene tags; nn is the number of val frames carrying the tag.

grouptagnnmAPNDSmATE
all–299526.540.20.636
areahighway196028.038.10.645
areacity32012.424.90.796
areaterminal32014.322.70.849
arearural31516.925.20.790
weatherclear156024.835.80.677
weatherovercast91528.841.00.620
weatherrain40019.528.00.795
lightingilluminated195529.442.10.629
lightingtwilight44020.127.40.719
lightingdark36021.928.30.787
lightingglare20016.124.30.804
daytimenoon151522.234.40.692
daytimemorning100034.642.30.668
daytimeevening48022.932.70.692

Table 17.

Table 18 crosses the same tags with range. The decline with range is present under every tag. Rain and glare cost a roughly constant fraction of the band’s mAP at every range (rain 21–35 %, glare 34–40% below the all-frames value of the band); city frames lose 42% at 0–25 m and about 72% beyond. The night-time and tunnel tags are absent or too small to report.

grouptag0–25 m mAP0–25 m NDS25–50 m mAP25–50 m NDS50–100 m mAP50–100 m NDS100–150 m mAP100–150 m NDS
all–43.047.025.339.416.830.59.419.2
areahighway42.543.535.143.220.631.510.619.1
areacity24.829.76.714.55.013.92.78.4
areaterminal27.030.512.820.04.310.72.75.7
arearural39.938.616.624.27.217.64.912.8
weatherclear40.045.421.531.716.227.211.319.8
weatherovercast48.750.338.645.720.033.09.718.7
weatherrain28.031.419.427.612.518.67.412.7
lightingilluminated46.449.234.844.117.332.510.519.8
lightingtwilight34.736.115.022.910.919.37.512.7
lightingdark26.831.222.429.022.628.69.514.8
lightingglare26.830.916.023.210.116.96.211.5
daytimenoon38.641.319.632.611.523.48.118.5
daytimemorning44.944.531.036.528.634.615.320.5
daytimeevening29.336.720.128.317.426.67.713.7

Table 18. EviScope (R50, seed 0), official protocol restricted to the val frames carrying each tag and to objects in each disjoint range band; mAP and NDS (%). Tags with at least 100 val frames.

Training trajectory

Figure 3(a) shows the training loss of the EVISCOPE and baseline runs (seed 0); the module changes neither the level nor the shape of the curve. Figure 3(b) repeats the probe of Table 5 on the model at intermediate iterations over the first 14,410 iterations (1.3 epochs). The correlation between local evidence and radius scale is ρ=−0.95\rho= -0.95 at the first point probed, iteration 2,280, and stays between −0.95-0.95 and −0.90-0.90 over the window; the final model’s value is −0.91-0.91 (§ 6.1). The range correlation stays positive and weaker throughout (0.21 to 0.67). The allocation is not a late-training artefact. Table 19 gives the val score at the end of epochs 1, 2 and 4 for seed 0 and for seed 2, the low seed of Table 7; multi-frame input is enabled from epoch 2. Seed 2 is behind from the first epoch (29.93 against 31.88 NDS) and never closes the gap at the reported checkpoints. These two trajectories do not establish when seed-dependent differences become fixed.

seed 0seed 0seed 0seed 2seed 2seed 2
epoch124124
mAP20.0323.7126.4717.8121.8225.77
NDS31.8836.5140.1929.9334.4137.73

Table 19. EviScope (R50) evaluated at the end of training epochs, seeds 0 and 2.

Figure 3. (a) Training loss, moving average over 40 iterations, epoch boundaries marked. (b) Spearman correlation of the emitted radius scale with local evidence and with range, probed on the EviScope model at intermediate iterations of the run.

Recall by range and evidence for EviScope

Table 20 reports the baseline. Table 21 gives the same cross-tabulation for EviScope, with the number of objects in each cell; Table 10 is the difference of the two. Recall rises in 16 of the 19 cells the table reports and the largest decrease is 1.1 points; the cell beyond 100 m with more than fifteen returns, omitted from the table, decreases by two objects out of 23.

Inference-time substitution, per run

Table 22 gives the absolute scores behind Table 6. Each block is one trained EviScope model, the first four ResNet-50 and the last V2-99; rows within a block share weights and differ only in the substitution applied at inference. Evaluations here use four GPUs rather than eight, which moves an absolute score by up to 0.09 mAP and 0.05 NDS against Table 7 and § 4.2; every difference within a block uses the same evaluation configuration. Table 23 gives the two control models of § 6.2; recall was not computed for them, and the mAP of the radial-mean row was not recorded.

substitutionrun 0run 2run 3run 4V2-99
mAPNDSrec.mAPNDSrec.mAPNDSrec.mAPNDSrec.mAPNDSrec.
none26.4840.1880.1025.7737.7679.5124.3837.3979.6425.7938.1179.2926.5537.9280.19
radial mean26.4340.0180.1525.8537.9879.6624.4937.2279.6925.6638.1279.2826.5637.8680.37
radial shuffled26.4440.0680.0225.7037.7579.5924.4336.9279.5025.7038.0479.2826.5337.7480.21
radial constant26.4540.1480.0625.6637.5579.5424.3537.4179.5125.7938.1879.4226.5737.8680.29
temporal uniform19.6031.9375.6613.6827.1171.0612.0724.9169.4011.3725.7069.5413.3227.4770.38

Table 22. Absolute mAP / NDS / recall (%) for every substitution and run.

substitutionno density head mAPno density head NDSevidence-blind control mAPevidence-blind control NDS
none26.4138.9126.0239.30
radial mean—38.81—39.25
temporal uniform19.4731.5717.3127.55

Table 23.

Radar degradation

Removing radar at evaluation time changes both the detector’s features and the evidence field, so it does not isolate the conditioner. Table 24 is the seed-0 pair re-scored on identical corrupted inputs. Thinning the returns, uniformly or preferentially at long range, leaves the within-pair difference between +0.55 and +0.70 mAP at every rate, including one that costs both models a fifth of their accuracy. Disabling two of the six radar sensors moves it to −0.28. In the light of Table 7 we do not read the sign of a single pair, and the similarity of the differences under thinning does not establish a robustness advantage. A possible limitation is that the descriptor contains no explicit sensor-coverage information: a zero count alone cannot distinguish missing coverage from an observed region with no returns. Whether this ambiguity explains the sector-failure result would require a separate control.

Qualitative examples

Figure 4 shows the radius scale αq\alpha_q that the trained module emits for the queries of the last decoder layer on three val frames. Queries that land on the carriageway, where returns are dense, contract (dark, small); queries in sparse regions beside the road receive relatively larger scales (light, large). The observed scales remain below the baseline value of one. The pattern is the same in clear, dusk and snow frames.

Bird’s-eye views of highway scenes with radar returns, ground-truth boxes, and decoded query centres

Figure 4. Bird’s-eye view; ego at the triangle, forward up, rings at 50, 100 and 150 m. Grey: radar returns. Black: ground-truth boxes. Discs: last-layer queries at their decoded centres, coloured and sized by the emitted radius scale αq\alpha_q.

Query-local evidence and object-level return counts

Section 3.1 counts radar returns inside ground-truth boxes, while the module reads the field at decoded query centres. To measure how far the second tracks the first, we hooked the last decoder iteration on 300 val samples spaced evenly over the split, read the five field channels at each query’s centre with the module’s own sampling function, and matched every ground-truth box within 150 m to Table 23: Absolute mAP / NDS for the two control models under inference-time substitution, seed 0. no density head: radial and temporal modulation trained without the auxiliary density loss. evidence-blind control: the early module that read the evidence field at raw polar coordinates rather than the decoded query centres (§ 4.2).

nearest unmatched query whose centre lies within 2 m (greedy, one-to-one). Boxes are taken before the evaluator’s point filter, so 47.2% have no annotated return, against 42.7% of evaluated objects in § 4.1. Table 25 reports two trained models.

rangeboxescoverageρ0\rho_0ρ1\rho_1ρ2\rho_2AUC1_1
0–150 m9,6840.855 / 0.8400.48 / 0.480.51 / 0.510.46 / 0.460.72 / 0.72
0–50 m4,0580.925 / 0.9140.45 / 0.450.47 / 0.470.38 / 0.380.68 / 0.67
50–100 m3,4580.820 / 0.8020.42 / 0.420.46 / 0.460.39 / 0.400.70 / 0.70
100–150 m2,1680.779 / 0.7630.33 / 0.350.42 / 0.430.40 / 0.410.72 / 0.72
returns in box01–23–56–15> 15
median one-cell reading15.421.936.751.4111.2
zero centre-cell reading20.2%8.1%2.4%0.9%0.0%
field built from, read atboxesρ1\rho_1zero readAUC1_1
keyframe sweep, box centre9,6840.5558.3%0.76
five sweeps (as used), box centre9,6840.5224.4%0.73
five sweeps (as used), matched query8,2770.5120.2%0.72

Table 25. Ground-truth return counts against the evidence read at the matched query’s centre, last decoder iteration, 300 val samples, two EviSCOPE models (seed 0 / seed 2). ρ\rho: Spearman correlation between the box’s return count and the log-count at dilation 0, 1, 2 cells. AUC: separation of zero-return from returning boxes by the one-cell reading. Coverage: share of boxes with a query centre within 2 m. Middle: median one-cell reading and share of zero centre-cell readings by the box’s return count, seed 0. Bottom: the non-learned field read at the true box centres, built from the keyframe sweep that the annotation counts or from the five sweeps per radar the module uses; zero read is the share of zero-return boxes reading below half a return in the centre cell.

The query-local reading rises monotonically with the object’s return count. Spearman correlation is 0.51 at the one-cell dilation and 0.42–0.47 within each range band, and the reading separates zero-return from returning objects with AUC 0.72 overall and 0.67–0.72 within bands; the two checkpoints agree to within 0.03. The correspondence is partial. Zero-return objects sit on a median one-cell reading of 15 returns, and only 20% of them read below half a return in the centre cell: the 2.4 m cell and its neighbourhood collect clutter and returns from adjacent objects. The descriptor therefore carries a graded, noisy proxy of an object’s evidence, not the zero-return condition used to motivate the method. Figure 5 shows the full distributions.

Figure 5. (a) Returns read at the matched query’s centre (one-cell dilation, log axis) against the returns annotated in the ground-truth box, by range band; boxes span the interquartile range, the bar is the median, whiskers the 5th–95th percentiles; model seed 0, 8,277 matched boxes; cells with fewer than 20 boxes are omitted. (b) ROC of the same reading for separating zero-return from returning boxes, per band; solid seed 0, dashed seed 2.

Where the gap comes from. The field has no parameters, so it can also be read at the ground-truth box centres, without a model and for all 9,684 boxes (bottom of Table 25). Built from the keyframe sweep alone, the sweep the annotation counts, the field reads below half a return in the centre cell for 58.3% of zero-return boxes; built as the module builds it, from the five most recent sweeps of each radar, for 24.4%; read at the matched query centre, for 20.2%. Rank correlation with the box’s count is 0.55, 0.52 and 0.51, and AUC 0.76, 0.73 and 0.72. The partial correspondence therefore comes mostly from how the field is built (five aggregated sweeps and 2.4 m cells read bilinearly) and only slightly from where a matched query sits relative to its box centre. As a check of the single-sweep reconstruction, every box with an annotated return has a keyframe return within half its diagonal. Whether a single-sweep field would condition sampling better is untested.

Table 24: Paired radar degradation, four-epoch models. Both models are scored on the same corrupted inputs: the dropout mask is seeded by sample token, so the comparison is per-point paired rather than per-run. Relative drop is measured against each model’s own clean score, so the two columns are not affected by the 0.65 mAP the module already holds when nothing is degraded. sector failure removes a third of the radar sensors.

Full-schedule control experiments

This appendix reports completed geometry-only and shuffled-evidence controls, a multi-seed density-head comparison, and two pairs at eight epochs. Each model is trained on one node of eight A100 GPUs with a per-GPU batch of one; cross-node evaluations are marked † in Table 26. Seeds 0–2 are the first three seeds of Table 7; they include the two highest-NDS EVISCOPE runs in that seven-seed set.

Table 26: Full-schedule controls; mAP / NDS (%). Superscripts identify results reproduced from other tables for paired reference (aTable 7; b seed-0 row +radius & window of Table 11; c seed-0 pair of Table 8). Only completed comparisons are reported; a dash denotes no reported pair. Δ\Delta is EVISCOPE minus the row’s model within a seed; means use the available pairs. \dagger Arms evaluated on different nodes (Appendix § 13); on the control’s node, EVISCOPE scores 26.50 / 40.19 (seed 0) and 26.59 / 40.29 (seed 1). \ddagger EVISCOPE with evidence inputs permuted at random across the queries of each sample in training and evaluation.

comparisonmodelseed 0seed 1seed 2mean Δ\DeltaNDS
four epochs
EVISCOPEa26.47 / 40.1926.51 / 40.2425.77 / 37.73
baselinea25.84 / 38.4826.48 / 38.8526.17 / 37.39
geometry-onlycontrol25.35 / 38.57\dagger25.47 / 37.98\dagger25.31 / 37.87+1.25
shuffled evidencecontrol\ddagger25.19 / 37.30\dagger26.31 / 38.25\dagger–+2.44
density headno density branch26.46 / 38.93b25.92 / 38.8126.03 / 38.42+0.67
eight epochs
EVISCOPE26.66 / 39.91c\dagger26.49 / 40.08–
baseline26.40 / 39.24c\dagger26.62 / 38.89–+0.93

Table 26.

Geometry-only conditioner. The module of § 4.2 with the five field channels of dqd_q replaced by zeros. It keeps the four geometric terms, the capacity, both learned branches, the identity initialisation and the auxiliary target computed from the true counts, so the only difference from EVISCOPE is whether the descriptor sees local evidence. Four epochs, seeds 0–2, paired with EVISCOPE and the baseline at the same seeds.

Density head across seeds. EVISCOPE against the same module trained without the density branch, four epochs.

Eight-epoch pairs. EVISCOPE against the baseline at eight epochs, with the completed seed-1 pair added to the seed-0 pair of Table 8. Both seed-1 arms are trained and evaluated on the same node, unlike the cross-node evaluation of the seed-0 pair.

Figure 6 places the completed main within-seed comparisons on one axis.

Figure 6. Completed within-seed comparisons, first model minus second. Dots are seeds; hollow dots mark arms evaluated on different nodes (†in Table 26). Diamonds and bars are mean ± SE; nn counts completed pairs. The four-epoch baseline and V2-99 comparisons, eight-epoch seed-0 pair and density-head seed-0 comparison are also reported in Table 7, Table 8 and Table 11, respectively.

configurationmAP↑NDS↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓
baseline, no module25.8438.480.6630.4170.5220.7110.132
+ radius & window26.4638.930.6620.4180.4900.7500.110
+ density head (EVISCOPE)26.4740.190.6360.4130.4730.6650.118
control: geometry-only (evidence inputs zeroed)25.3538.570.6750.4280.4900.7030.115
control: radius fixed at 0.6×, no module25.6938.020.6600.4260.5030.7800.114
same, seed 125.4537.630.6760.4340.5490.7150.136

Table 11. Component ablation at the full four-epoch schedule, seed 0, single runs. The first three rows are cumulative and share every other setting. The last two rows are the baseline with every query’s sampling extent fixed at 0.6×, the value the trained module most often emits, and no module. The geometry-only control is EVISCOPE with its five evidence inputs set to zero (Appendix § 30).

Geometry-only result. Over seeds 0–2, EVISCOPE exceeds the geometry-only control by +0.87±0.21+0.87 \pm0.21 mAP and +1.25±0.72+1.25 \pm0.72 NDS (mean ± SE; Table 26). mAP is higher in all three pairs (+1.12+1.12, +1.04+1.04, +0.46+0.46) and translation error is lower in all three (mATE 0.675 →\to 0.636, 0.657 →\to 0.624, 0.662 →\to 0.648). The NDS differences are +1.62+1.62, +2.26+2.26 and −0.14-0.14. Orientation and velocity error are lower at seeds 0 and 1 and higher at seed 2 (mAOE 0.490 →\to 0.473, 0.510 →\to 0.499, 0.526 →\to 0.545; mAVE 0.703 →\to 0.665, 0.751 →\to 0.641, 0.744 →\to 0.768); these two terms account for the sign of the seed-2 NDS difference. The observed control mean NDS is 38.14 against the baseline’s 38.24 and EviSCOPE’s 39.39, and its mAP is below the baseline’s in all three pairs (−0.49-0.49, −1.01-1.01, −0.86-0.86). Re-evaluated on the control’s node across checkpoints, the three-seed difference progresses from +0.11±0.46+0.11 \pm0.46 mAP / +0.20±0.76+0.20 \pm0.76 NDS at epoch 2 to +0.33±0.15+0.33 \pm0.15 / +0.55±0.50+0.55 \pm0.50 at epoch 3 and +0.91±0.22+0.91 \pm0.22 / +1.26±0.73+1.26 \pm0.73 at epoch 4. The per-seed NDS ordering is not uniform: differences for seeds 0–2 are (+1.66+1.66, −0.18-0.18, −0.87-0.87) at epoch 2, (<−0.01< -0.01, +1.55+1.55, +0.10+0.10) at epoch 3, and (+1.62, +2.31, −0.14) at epoch 4. Across the 75 validation scenes, 1,000× scene bootstrap on the three-seed mean gives 95% confidence intervals of [+0.23, +1.72] mAP and [−0.28, +2.48] NDS ([−0.08, +2.38] mAP and [−1.05, +3.51] NDS under two-level scene-and-seed bootstrap over n=3n=3 seeds). In 1,000 random scene splits, the seed-mean Δ\DeltaNDS is positive in 94.2% of halves and in both halves of 88.3% of splits. In the 100–150 m band, seeds 0–2 give Δ\DeltamAP of +0.7, +0.1, −0.5 and Δ\DeltaNDS of +1.1, +1.2, −0.3. Thus the NDS intervals include zero, and neither the checkpoint nor the far-range ordering holds uniformly across seeds.

By scene condition. Table 27 splits each pair by the official TruckScenes scene tags, keeping the 20 conditions with at least 300 validation frames, and scores each subset with the official metrics on its frames. The seed-mean difference is positive in 19 of the 20 conditions; the exception is rural scenes (−0.11 NDS).

Table 27: EVISCOPE minus the geometry-only control by official scene condition, Δ\DeltaNDS (points) at every seed. Conditions with at least 300 of the 2,995 validation frames; each frame carries one tag per category, so rows of different categories overlap. Each subset is scored with the official metrics restricted to its frames. Seed-0 and seed-1 EVISCOPE predictions are those of Table 26.

conditionframesseed 0seed 1seed 2mean
areacity320+1.38+3.47+0.92+1.92
highway1,960+0.11+2.46+1.22+1.26
rural315+0.46+0.89−1.68−0.11
terminal320−0.13+0.83−0.21+0.16
weatherclear1,560+1.82+2.50−1.72+0.87
overcast915−1.03+1.85+1.22+0.68
rain400−0.37+0.50+0.35+0.16
lightingdark360+2.01−0.48−0.14+0.46
illuminated1,955+0.89+2.54+0.09+1.17
twilight440+1.72+0.44+0.59+0.92
daytimeevening480+3.92−0.99+0.67+1.20
morning1,000+0.22+3.70−1.26+0.89
noon1,515+1.32+2.45−0.45+1.11
seasonautumn1,360+2.05+2.50+0.23+1.59
summer880+0.38+3.22−0.95+0.88
winter755+1.28+1.15+0.55+0.99
structureoverpass360+1.65+3.14−0.06+1.58
regular1,995+0.61+4.37+0.30+1.76
underpass480−3.85+1.72+2.34+0.07
constructionunchanged2,875+1.63+2.18−0.16+1.22

Table 27.

What the control learns. We ran the probe of Table 5 (last decoder iteration, first 200 validation samples, 36,000 queries per model) on the three control models and on EVISCOPE seeds 0–4, and computed rank correlations within each range band. EVISCOPE’s radius falls as the evidence at the query rises, in every band and seed (Spearman −0.89 to −0.95), and its effective window shortens (−0.06 to −0.73, negative in all 20 seed-by-band cells). The control’s radius stays at the branch’s lower bound, 0.600–0.607 for every query (standard deviation below 0.001 at each seed), and its window has no consistent relation to the evidence (−0.26 to +0.37). Without the evidence input, every query samples as narrowly as EVISCOPE samples its best-observed queries (Figure 7).

Figure 7. Allocation with and without the evidence input, inside each range band. Probe of Table 5 on EVISCOPE (blue, filled) and the geometry-only control (grey, hollow), seeds 0–2 of each. Queries are binned by the evidence read at the query: the first bin holds zero readings, the other eight are quantile bins of the positive readings pooled over all models. Top: mean radius scale αq\alpha_q; bottom: mean effective window exp⁡(H)\exp(\mathrm{H}) of the frame weights, in frames. Bins with fewer than 50 queries are omitted. The control’s radius is 0.600–0.607 for every query.

Eight-epoch result. The seed-1 pair, trained and evaluated on one node, gives EVISCOPE 26.49 / 40.08 and the baseline 26.62 / 38.89: +1.19 NDS and −0.13 mAP. With the seed-0 pair of Table 8 (+0.67 NDS, +0.26 mAP), EVISCOPE is above its baseline on NDS in both completed eight-epoch pairs, by +0.93 ±\pm 0.26 on average (mean ±\pm SE), with mAP level (+0.07 ±\pm 0.20). At seed 1 the NDS difference comes from orientation, velocity and scale error (mAOE 0.548 →\to 0.457, mAVE 0.743 →\to 0.719, mASE 0.420 →\to 0.407); translation error is similar (0.603 and 0.602). These two pairs do not establish an improvement with longer training.

Density-head result. Over seeds 0–2, EVISCOPE exceeds the same module trained without the density branch by +0.67 ±\pm 0.68 NDS and +0.11 ±\pm 0.25 mAP (mean ±\pm SE). The NDS differences are +1.26, +1.43 and −0.69, and the mAP differences +0.01, +0.59 and −0.26. Against the baseline, the no-density variant averages +0.48 ±\pm 0.31 NDS and −0.03 ±\pm 0.35 mAP; it is also above the geometry-only control in all three pairs (mAP +1.11, +0.45, +0.72; NDS +0.36, +0.83, +0.55; evaluated on different nodes at seeds 0 and 1), noting that those two models differ in both the evidence input and the auxiliary loss.

Shuffled-evidence result. Permuting the evidence channels across the queries of each sample gives 25.19 / 37.30 at seed 0 and 26.31 / 38.25 at seed 1 (mAP / NDS), so EVISCOPE exceeds the shuffled control in both completed pairs: +1.28 mAP / +2.89 NDS at seed 0 and +0.20 mAP / +1.99 NDS at seed 1, averaging +0.74 ±\pm 0.54 mAP and +2.44 ±\pm 0.45 NDS (+0.80 ±\pm 0.52 mAP and +2.47 ±\pm 0.43 NDS when evaluated on the control’s node). Translation, orientation and velocity errors are higher in the shuffled control at both seeds (mATE 0.664 vs. 0.636 and 0.652 vs. 0.624; mAOE 0.578 vs. 0.473 and 0.567 vs. 0.499; mAVE 0.740 vs. 0.665 and 0.720 vs. 0.641), and both shuffled runs also fall below their matched baselines (25.84 / 38.48 and 26.48 / 38.85). The ordering against geometry-only is mixed: shuffled minus geometry-only is −0.16 / −1.27 at seed 0 and +0.84 / +0.27 at seed 1 (mAP / NDS). These two pairs support the value of spatial alignment under this control, without establishing an ordering between all control variants.

Settings of every comparison

Table 28 lists, for every comparison in the paper, the schedule, the number of runs and what differs between its rows. All training runs use one node of eight A100 GPUs with a per-GPU batch of one, ResNet-50 at 256×704256 \times704 unless stated, and the official TruckScenes evaluator.

References

  1. [1]Geonho Bang, Geunju Baek, Dongyoung Lee, Wonjun Jeong, and Jun Won Choi. Horizon3D: Sparse radar-camera fusion for long-range 3d perception in autonomous driving. In European Conference on Computer Vision (ECCV), 2026.
  2. [2]Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020.
  3. [3]Dallas Card, Peter Henderson, Urvashi Khandelwal, Robin Jia, Kyle Mahowald, and Dan Jurafsky. With little power comes great responsibility. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020.arxiv.org/abs/2010.06595
  4. [4]Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), 2020.arxiv.org/abs/2005.12872
  5. [5]Xiaomeng Chu, Jiajun Deng, Guoliang You, Yifan Duan, Houqiang Li, and Yanyong Zhang. RaC-Former: Towards high-quality 3d object detection via query-based radar-camera fusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025.
  6. [6]Jesse Dodge, Gabriel Ilharco, Roy Schwartz, Ali Farhadi, Hannaneh Hajishirzi, and Noah Smith. Fine-tuning pretrained language models: Weight initializations, data orders, and early stopping. arXiv preprint arXiv:2002.06305, 2020.
  7. [7]Felix Fent, Fabian Kuttnerich, Florian Ruch, Farija Rizwin, Stefan Juergens, Lorenz Lechermann, Christian Nissler, Andrea Perl, Ulrich Voll, Min Yan, and Markus Lienkamp. MAN TruckScenes: A multimodal dataset for autonomous trucking in diverse conditions. In Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track, 2024.
  8. [8]Andrew Gelman and Eric Loken. The statistical crisis in science. American Scientist, 102:460, 2014.
  9. [9]Zijian Gu, Jianwei Ma, Yan Huang, Honghao Wei, Zhanye Chen, Hui Zhang, and Wei Hong. HGSFusion: Radar-camera fusion with hybrid generation and synchronization for 3D object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pp. 3185–3193, 2025.
  10. [10]Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016.arxiv.org/abs/1512.03385
  11. [11]Geoffrey E. Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, and Ruslan R. Salakhutdinov. Improving neural networks by preventing co-adaptation of feature detectors. arXiv preprint arXiv:1207.0580, 2012.
  12. [12]Sara Hooker, Dumitru Erhan, Pieter-Jan Kindermans, and Been Kim. A benchmark for interpretability methods in deep neural networks. In Advances in Neural Information Processing Systems (NeurIPS), 2019.
  13. [13]Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022.
  14. [14]Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021.
  15. [15]Xun Huang, Ziyu Xu, Hai Wu, Jinlong Wang, Qiming Xia, Yan Xia, Jonathan Li, Kyle Gao, Chenglu Wen, and Cheng Wang. L4DR: Lidar-4dradar fusion for weather-robust 3d object detection. In AAAI Conference on Artificial Intelligence, 2025.
  16. [16]Xiaohui Jiang, Shuailin Li, Yingfei Liu, Shihao Wang, Fan Jia, Tiancai Wang, Lijin Han, and Xiangyu Zhang. Far3D: Expanding the horizon for surround-view 3d object detection. In AAAI Conference on Artificial Intelligence, 2024.
  17. [17]Jisong Kim, Minjae Seong, and Jun Won Choi. CRT-Fusion: Camera, radar, temporal fusion using motion information for 3d object detection. In Advances in Neural Information Processing Systems (NeurIPS), 2024.arxiv.org/abs/2411.03013
  18. [18]Youngseok Kim, Sanmin Kim, Jun Won Choi, and Dongsuk Kum. Craft: Camera-radar 3d object detection with spatio-contextual fusion transformer. In AAAI Conference on Artificial Intelligence, 2023a.
  19. [19]Youngseok Kim, Juyeb Shin, Sanmin Kim, In-Jae Lee, Jun Won Choi, and Dongsuk Kum. Crn: Camera radar net for accurate, robust, efficient 3d perception. In IEEE/CVF International Conference on Computer Vision (ICCV), 2023b.arxiv.org/abs/2304.00670
  20. [20]Yiheng Li, Yang Yang, and Zhen Lei. RCTrans: Radar-camera transformer via radar densifier and sequential decoder for 3d object detection. In AAAI Conference on Artificial Intelligence, 2025.arxiv.org/abs/2412.12799
  21. [21]Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In AAAI Conference on Artificial Intelligence, 2023.
  22. [22]Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird’s-eye-view representation from multi-camera images via spatiotemporal transformers. In European Conference on Computer Vision (ECCV), 2022.arxiv.org/abs/2203.17270
  23. [23]Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d v2: Recurrent temporal fusion with sparse model. arXiv preprint arXiv:2305.14018, 2023.arxiv.org/abs/2305.14018
  24. [24]Zhiwei Lin, Zhe Liu, Yongtao Wang, Le Zhang, and Ce Zhu. RCBEVDet++: Toward high-accuracy radar-camera fusion 3D perception network. arXiv preprint arXiv:2409.04979, 2024a.arxiv.org/abs/2409.04979
  25. [25]Zhiwei Lin, Zhe Liu, Zhongyu Xia, Xinhao Wang, Yongtao Wang, Shengxiang Qi, Yang Dong, Nan Dong, Le Zhang, and Ce Zhu. Rcbevdet: Radar-camera fusion in bird’s eye view for 3d object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024b.arxiv.org/abs/2403.16440
  26. [26]Haisong Liu, Yao Teng, Tao Lu, Haiguang Wang, and Limin Wang. Sparsebev: High-performance sparse 3d object detection from multi-camera videos. In IEEE/CVF International Conference on Computer Vision (ICCV), 2023a.arxiv.org/abs/2308.09244
  27. [27]Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETR: Position embedding transformation for multi-view 3d object detection. In European Conference on Computer Vision (ECCV), 2022.arxiv.org/abs/2203.05625
  28. [28]Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. BEVFusion: Multi-task multi-sensor fusion with unified bird’s-eye view representation. In IEEE International Conference on Robotics and Automation (ICRA), 2023b.arxiv.org/abs/2205.13542
  29. [29]Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the value of network pruning. In International Conference on Learning Representations (ICLR), 2019.arxiv.org/abs/1810.05270
  30. [30]Yunfei Long, Abhinav Kumar, Xiaoming Liu, and Daniel Morris. RICCARDO: Radar hit prediction and convolution for camera-radar 3d object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025.arxiv.org/abs/2504.09086
  31. [33]Ramin Nabati and Hairong Qi. Centerfusion: Center-based radar and camera fusion for 3d object detection. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2021.DOI
  32. [35]Dong-Hee Paek, Seung-Hyun Kong, and Kevin Tirta Wijaya. K-radar: 4d radar object detection for autonomous driving in various weather conditions. In Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track, 2022.
  33. [36]Andras Palffy, Ewoud Pool, Srimannarayana Baratam, Julian F. P. Kooij, and Dariu M. Gavrila. Multi-class road user detection with 3+1d radar in the view-of-delft dataset. IEEE Robotics and Automation Letters, 7:4961–4968, 2022.DOI
  34. [37]Heejun Park, Jaesok Jeong, and Kuk-Jin Yoon. RAF: Reliability-aware fusion of camera, lidar, and 4d radar for robust 3d object detection in adverse weather. In European Conference on Computer Vision (ECCV), 2026.arxiv.org/abs/2607.04587
  35. [38]Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. In International Conference on Learning Representations (ICLR), 2023.arxiv.org/abs/2210.02443
  36. [39]Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In European Conference on Computer Vision (ECCV), 2020.
  37. [40]David Picard. torch.manual_seed(3407) is all you need: On the influence of random seeds in deep learning architectures for computer vision. arXiv preprint arXiv:2109.08203, 2021.arxiv.org/abs/2109.08203
  38. [41]Joseph P. Simmons, Leif D. Nelson, and Uri Simonsohn. False-positive psychology: Undisclosed flexibility in data collection and analysis allows presenting anything as significant. Psychological Science, 22:1359–1366, 2011.
  39. [43]Ruochen Wang, Minhao Cheng, Xiangning Chen, Xiaocheng Tang, and Cho-Jui Hsieh. Rethinking architecture selection in differentiable nas. In International Conference on Learning Representations (ICLR), 2021a.arxiv.org/abs/2108.04392
  40. [44]Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xiangyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In IEEE/CVF International Conference on Computer Vision (ICCV), 2023.arxiv.org/abs/2303.11926
  41. [45]Yue Wang, Vitor Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning (CoRL), 2021b.arxiv.org/abs/2110.06922
  42. [46]Philipp Wolters, Johannes Gilg, Torben Teepe, Fabian Herzog, Felix Fent, and Gerhard Rigoll. SpaRC: Sparse radar-camera fusion for 3d object detection. arXiv preprint arXiv:2411.19860, 2024.arxiv.org/abs/2411.19860
  43. [47]Philipp Wolters, Johannes Gilg, Torben Teepe, Fabian Herzog, Anouar Laouichi, Martin Hofmann, and Gerhard Rigoll. Unleashing hydra: Hybrid fusion, depth consistency and radar for unified 3d perception. In IEEE International Conference on Robotics and Automation (ICRA), 2025.arxiv.org/abs/2403.07746
  44. [48]Tianwei Yin, Xingyi Zhou, and Philipp Krähenbühl. Center-based 3d object detection and tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021.arxiv.org/abs/2006.11275
  45. [49]Jingtong Yue, Zhiwei Lin, Xin Lin, Xiaoyu Zhou, Xiangtai Li, Lu Qi, Yongtao Wang, and Ming-Hsuan Yang. RobuRCDet: Enhancing robustness of radar-camera fusion in bird’s eye view for 3d object detection. In International Conference on Learning Representations (ICLR), 2025.arxiv.org/abs/2502.13071
  46. [51]Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492, 2019.arxiv.org/abs/1908.09492
  47. [52]Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations (ICLR), 2021.arxiv.org/abs/2010.04159

Paper details

vlnc:W_zdyu9r62
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

generative AI tools

View statement
Looking for related papers…
Contents