WorksComputer Science
HIVE-ECIA: Identity-Aware Historical Evidence Acquisition for Bandwidth-Efficient Multi-Object Tracking
Abstract
Panoramic multi-object tracking (MOT) requires reliable identity association across a 360° field of view, yet bandwidth-constrained operation forces a trade-off between low-resolution (LR) identity ambiguity and costly high-resolution (HR) transmission. We propose HIVE-ECIA, an identity-aware framework that selectively retrieves historical HR evidence to recover fragmented identities without modifying detection boxes or the underlying tracker. An ECIA Base-Guided query policy determines when additional HR evidence is warranted; HIVE ranks historical LR observations to select an informative retrieval timestamp; and ECIA Full integrates the retrieved HR crop through learned gated fusion and conservative identity safeguards. We further introduce DualPano-MOT, a real-world dual-fisheye panoramic evaluation dataset featuring controlled pedestrian interactions, panorama-seam challenges, and occlusions, with model-assisted, manually verified tracking annotations. On held-out QuadTrack sequences and zero-shot DualPano-MOT, HIVE-ECIA improves identity association across four MOT trackers without tracker-specific retraining. Across these trackers, identity switches decrease by up to 21.8% and 23.4%, while the largest IDF1 gains reach 3.81 and 6.74 percentage points, respectively. Under an offline JPEG-based acquisition model, our framework reduces encoded data volume by 89.2% and 79.3% relative to continuous Full-HR acquisition.
Introduction
Panoramic cameras provide a field of view (FoV), enabling comprehensive scene perception for robotics, navigation, and multi-object tracking (MOT). However, distributing resolution across such a wide FoV limits discriminative appearance details, particularly for small or distant pedestrians. Geometric distortion and occlusion further challenge identity association [17]. These limitations become critical in bandwidth-constrained edge–server systems: low-resolution (LR) panoramas reduce transmission cost but may cause identity fragmentation, whereas continuously transmitting high-resolution (HR) panoramas incurs substantial communication overhead. As illustrated in Figure 1, we address this trade-off by selectively acquiring historical HR evidence to recover ambiguous identities.

Figure 1. Overview of HIVE–ECIA. (a) DualPano-MOT panoramic acquisition. (b) Identity-association gains and HR data savings. (c) Historical HR crop retrieval enables identity recovery without modifying tracking geometry.
Existing MOT methods, including SORT [3], ByteTrack [27], OC-SORT [6], and BoT-SORT [1], improve identity association through motion, detection confidence, and appearance cues. Memory-based trackers [5, 9] further exploit historical representations, while OmniTrack [17] addresses panoramic distortions and tracking uncertainty. However, these approaches primarily reason over available observations, rather than determining what additional visual evidence should be acquired when identity information is insufficient.
Adaptive video analytics [8, 26] addresses communication constraints through selective visual fidelity, but identity recovery presents a different acquisition objective: the most recent or highest-resolution observation is not necessarily the most discriminative for resolving an ambiguous identity. Moreover, additional evidence may reinforce an incorrect association if incorporated without assessing its reliability. This exposes a gap between identity association and identity-aware evidence acquisition. Addressing it requires jointly deciding when additional HR evidence is necessary, which historical observation to retrieve, and how to integrate that evidence into reliable identity recovery.
We propose HIVE–ECIA, an identity-aware historical evidence acquisition framework for bandwidth-efficient panoramic MOT. Operating on a continuous LR stream, HIVE–ECIA selectively retrieves historical HR person crops from a timestamp-indexed buffer. The ECIA Base-Guided query policy first evaluates competing identity hypotheses to merge directly, request HR evidence, or abstain. When retrieval is needed, HIVE ranks eligible historical LR observations to select an informative HR timestamp. Finally, ECIA Full incorporates the retrieved evidence through learned residual gain and gated fusion, followed by safety-aware reassociation. Identity corrections update canonical IDs without modifying detections or tracking geometry, enabling integration with different MOT trackers.
We further introduce DualPano-MOT, a stationary dual-fisheye panoramic evaluation collection complementing the mobile-platform QuadTrack benchmark. It contains five sequences featuring pedestrian crossings, occlusions, and panorama-seam interactions, with 4,969 annotated frames and 8,518 person bounding boxes. We evaluate HIVE–ECIA across four MOT trackers on both datasets.
Our main contributions are summarized as follows:
We introduce HIVE–ECIA, an identity-aware historical evidence acquisition framework that jointly addresses when, which, and how HR evidence should be acquired and integrated for identity recovery. It combines Base-Guided acquisition, learned historical evidence ranking, and evidence-conditioned reassociation while preserving tracking geometry.
We present DualPano-MOT, a real-world dual-fisheye panoramic MOT evaluation collection with model-assisted, manually verified annotations. Its five sequences provide complementary identity-ambiguity scenarios involving pedestrian crossings, occlusions, and panorama-seam interactions.
Experiments across four MOT trackers demonstrate consistent identity-association improvements without tracker-specific retraining. HIVE–ECIA reduces identity switches by up to 21.8% on QuadTrack and 23.4% on DualPano-MOT. Under an offline JPEG-based acquisition model, it reduces encoded data volume by 89.2% and 79.3%, respectively, relative to continuous Full-HR acquisition.
Related Work
Panoramic visual tracking. Panoramic imagery provides comprehensive surrounding perception but introduces geometric distortion, resolution dilution, and boundary discontinuities. The 360VOT benchmark investigates these challenges in omnidirectional visual tracking [11], while JRDB provides panoramic observations for robotic human tracking [19]. OmniTrack [17] addresses panoramic MOT through distortion-aware feature modeling and trajectory feedback, introducing the QuadTrack benchmark. OmniTrack++ [16] further incorporates ExpertTrack Memory for long-term appearance consolidation and identity association. Unlike these approaches, which primarily exploit acquired panoramic observations and historical representations, HIVE–ECIA selectively retrieves additional historical high-resolution (HR) evidence to resolve identity ambiguity under constrained transmission budgets.
Multi-object tracking and identity association. Tracking-by-detection methods associate frame-wise detections using motion, geometry, and appearance cues. SORT [3] provides a lightweight motion-based baseline, while ByteTrack [27] incorporates low-confidence detections. OC-SORT [6] and MotionTrack [22] improve motion modeling, whereas DeepSORT and BoT-SORT [1, 24] exploit appearance representations. QDTrack [21] learns quasi-dense instance similarities, while Deep OC-SORT [18] adaptively integrates ReID into motion-based association. End-to-end methods such as TrackFormer, MOTR, and MOTRv2 [20, 25, 28] propagate learned tracking representations, while MOTIP [10] formulates association as identity prediction. Memory-based approaches, including MeMOT and MeMOTR [5, 9], maintain historical identity representations. In contrast, HIVE–ECIA actively selects previously untransmitted HR observations and incorporates them through guarded identity-recovery decisions.
Bandwidth-aware visual analytics. Resource-efficient video analytics balances inference quality and system cost. AWStream [26] adapts video transmission settings, Chameleon [12] optimizes inference configurations, and Reducto [13] filters redundant frames. DDS [8] uses server-side feedback to request higher-quality image regions, demonstrating selective visual acquisition. However, these systems primarily optimize general inference efficiency rather than identity consistency. HIVE–ECIA instead conditions HR acquisition on competing identity hypotheses, ranks historical observations using LR evidence, and employs gated fusion with conservative merge-or-abstain decisions. This couples bandwidth-aware acquisition with identity recovery.
Proposed Method
Framework Overview
We propose HIVE–ECIA, an identity-aware historical evidence acquisition framework for bandwidth-efficient multi-object tracking (MOT). As illustrated in Figure 2, our framework follows an edge–server design that continuously transmits low-resolution (LR) panoramas while retaining high-resolution (HR) observations at the edge for selective historical retrieval.

Figure 2. Overview of HIVE–ECIA. Given a continuous low-resolution (LR) panoramic stream, our framework determines when high-resolution (HR) evidence is needed, which historical observation to retrieve, and how to incorporate it for identity recovery. A safety-aware identity registry corrects track identities while preserving the original tracking geometry.
Edge–server acquisition. At the edge, captured HR panoramas are maintained in a timestamp-indexed buffer , while their downsampled LR counterparts are transmitted to the server. Given an LR frame at time , a detector produces detections , which are processed by a tracker to obtain raw tracks:
where and denote the raw tracks and tracker state, respectively. Track lifecycle, motion, and available appearance cues generate feasible Lost-to-New and Active-to-Active association candidates without ground-truth identity information.
For each candidate, let denote the proposed predecessor identity, the current target track, and the competing identities. The hypothesis set is . A frozen appearance encoder extracts embeddings from LR crops. For each hypothesis , an identity prototype is constructed from its nonempty LR observation history :
where denotes normalization.
Identity-aware HR retrieval. HIVE–ECIA follows a three-stage When–Which–How recovery pipeline. ECIA Base first determines whether an LR association can be resolved directly or requires additional HR evidence. For triggered queries, HIVE selects a historical timestamp from eligible LR observations, and the server retrieves at most one corresponding HR person crop from the edge-side buffer. ECIA Full then integrates the retrieved evidence for safety-aware identity reassociation. Accepted corrections update only canonical output identities, leaving detections, tracker states, and tracking geometry unchanged. A detailed module-level architecture is provided in Supplementary Fig. 6.
ECIA Base-Guided HR Query
The first component determines when additional HR evidence should be acquired. Rather than querying HR for every ambiguous association, a frozen ECIA Base model evaluates competing identity hypotheses using LR evidence.
For each proposed association , the predecessor is selected using tracker lifecycle, motion, and appearance cues, while competitors are drawn from other recently observed tracks. These are candidate hypotheses, not ground-truth identity assignments.
Let denote the appearance embedding of the current LR target crop . Given the identity prototypes defined in Eq. (2), a shared projection produces and . For each hypothesis , we construct the pairwise representation
where denotes element-wise multiplication. A lightweight MLP predicts the Base association logit .
ECIA Base computes two probability distributions:
where is a learned abstention logit. The identity-only distribution measures competition among identity hypotheses, whereas additionally permits abstention.
Let denote the complete decision set. We quantify association uncertainty using the predecessor decision margin, competitor margin, and normalized identity entropy:
The decision margin compares the proposed predecessor against all alternatives, including abstention, whereas the competitor margin considers only alternative identity hypotheses. Both margins are computed in probability space for the HR query policy, distinct from the logit-space margin used internally by the frozen ECIA Base model. When , the maximum competitor probability is defined as zero; in this case, the competitor margin provides no discriminative information between identities, and the decision margin still accounts for abstention.
The Base-Guided policy selects among three actions:
The direct-merge condition requires to rank first in , with and .
The HR request condition requires to rank first among identity hypotheses, , and at least one uncertainty signal:
The direct-merge condition takes precedence over HR requests; otherwise, the policy abstains when neither condition is satisfied. All thresholds are fixed during evaluation. Before HR retrieval, a deterministic pre-query safety check rejects infeasible identity merges under the current registry, avoiding unnecessary HR acquisition.
HIVE: LR Historical Candidate Ranking
When in Eq. (6) and pre-query safety is satisfied, HIVE selects a historical observation for HR retrieval. A bounded LR candidate bank is constructed from previously observed LR crops (), using appearance similarity, recency, and temporal diversity. Only causally available observations with valid historical HR correspondences are eligible.
HIVE employs a lightweight appearance encoder to extract LR features. Specifically, represents the current target crop, while aggregates the predecessor’s LR history using masked attention. The optional target-history representation summarizes earlier observations of the newly initialized track; a learned null embedding is used when this history is empty. Each historical candidate is represented by . The temporal metadata encodes the time gap through normalized linear, squared, and logarithmic transformations. We refer to as the query context, which is shared across candidates, while provides candidate-specific information.
HIVE jointly evaluates these features:
where denotes the learned identity-relevance score of historical candidate . Ranking relies exclusively on available LR observations, without accessing HR content or future frames. If , HR retrieval is skipped and the original identity assignment is retained.
HIVE is trained using class-weighted binary cross-entropy and event-normalized pairwise logistic ranking [4]:
where and denote positive and negative candidates for event . Only events containing both classes contribute to , weighted by .
The selected candidate determines the historical timestamp and LR region whose corresponding HR crop is retrieved using the LR–HR coordinate mapping. A frozen appearance encoder then extracts
The resulting HR embedding is passed to ECIA Full for identity reassociation. At most one HR crop is retrieved per query, without full-frame HR detection. If retrieval fails, HR-based recovery is skipped.
ECIA Full: HR Gain and Gated Fusion
Given the selected HR embedding , ECIA Full determines how HR evidence influences identity recovery. The frozen shared projection produces . Let . We compute the HR identity distribution and its Jensen–Shannon [14] disagreement with the LR Base distribution:
where is the HR temperature and is the mixture probability of identity hypothesis .
For each hypothesis , we construct an HR-conditioned relation:
where includes HR–LR similarities, Base association statistics, and the JS disagreement . The event-level feature vector combines Base confidence, identity margins, HR similarities, and with additional cross-resolution disagreement cues.
Two lightweight MLPs, and , predict a per-hypothesis signed HR residual and an event-level gate, respectively. Their outputs are fused with the frozen Base logits:
where denotes the sigmoid function. The signed residual can strengthen or suppress each identity hypothesis, while the shared gate controls the influence of HR evidence.
The final identity distribution is
where is inherited from the frozen Base. A merge requires to rank first with sufficient confidence and decision margin; otherwise, ECIA abstains.
Recovery-aware training. ECIA Base is trained first and frozen while HR Gain and Gate are optimized:
The terms enforce classification, margin separation, merge correctness, Base preservation, HR-assisted recovery, and gate supervision, respectively. In particular, preservation protects correct Base decisions, while recovery emphasizes errors correctable by verified HR evidence. Ground truth is used only for offline supervision.
Safe identity recovery. All merge proposals undergo Final Safety before updating the canonical identity registry . For an accepted merge,
where is a generation-safe raw track token, with raw ID and reuse generation . The registry prevents active-ID collisions and raw-ID reuse ambiguity. Corrections affect only current and future output IDs, preserving tracking geometry, internal tracker states, and past predictions.
DualPano-MOT Dataset
We introduce DualPano-MOT, a real-world dual-fisheye panoramic MOT evaluation dataset designed to study identity recovery under controlled pedestrian interactions.
Data collection. Two opposing fisheye cameras mounted on a stationary rig capture complementary views. Timestamp-aligned frames are dewarped and stitched into equirectangular panoramas, preserving source-frame correspondences. Original-resolution panoramas serve as HR references, while LR observations are generated by spatial downsampling for tracking and selective historical retrieval.
Identity-oriented scenarios. DualPano-MOT comprises five controlled scenarios: normal walking, pedestrian crossing, panorama-seam interaction, partial occlusion, and full occlusion. Tracking annotations are automatically initialized and manually verified in CVAT to preserve sequence-consistent identities. The collection contains 6,969 stitched frames, including 4,969 annotated frames with 8,518 person bounding boxes and 10 sequence-local identity trajectories.
Dataset characteristics. As shown in Figure 3(a), 64.6% of annotated objects occupy less than of the panorama image area, with below , highlighting limited target appearance detail. Moreover, the annotations contain 74 within-track observation gaps. Although only nine gaps span at least 30 frames, they account for 73.8% of absent identity-frames (Figure 3(b)). These gaps quantify temporal observation discontinuity rather than identity switches.

Figure 3. DualPano-MOT characteristics. (a) Cumulative distribution of annotated object sizes relative to panorama area. (b) Short ( frames) and long ( frames) observation gaps, measured by event counts and absent identity-frames.
As summarized in Table 1, DualPano-MOT complements existing panoramic benchmarks by combining staged identity-recovery scenarios, explicit seam interactions, and traceable dual-fisheye observations in a controlled evaluation setting.
Table 1. Comparison of tracking datasets. Pano.: panoramic imagery; MOT: multi-object identity annotations; Staged: controlled identity-recovery scenarios; Border: explicit panorama-border or seam challenges; Trace: source-view-to-panorama correspondence. “–” denotes an attribute not established by the cited benchmark description.
Experiments
Experimental Setup
Datasets and protocol. We evaluate HIVE–ECIA on QuadTrack [17] (2,400 frames, four held-out sequences) and DualPano-MOT (§ 4, 4,969 annotated frames). DualPano-MOT is evaluated zero-shot, without dataset-specific training or tuning. Panoramas are downsampled to (QuadTrack) and (DualPano-MOT) before detection and tracking, while original-resolution frames are retained for selective historical HR retrieval.
Implementation. Experiments are conducted on an NVIDIA RTX 5090 GPU. We use YOLO11l with SORT, ByteTrack, OC-SORT, and BoT-SORT, together with a frozen OSNet-AIN [29] appearance encoder. HIVE and ECIA checkpoints are fixed across tracker families. Each baseline–recovery pair shares identical detections, and recovery modifies only output identities.
Evaluation. We report HOTA and AssA [15], IDF1 [23], and IDSW [2]. Bandwidth is estimated using JPEG-encoded payloads (quality 90), rather than measured network transmission. Ground truth is excluded from online inference. Additional details are provided in the supplementary.
Comparison with Tracking Baselines
Table 2 demonstrates consistent identity-association improvements across four tracker families and two datasets. HOTA, IDF1, and AssA increase in all eight comparisons, accompanied by fewer identity switches, whereas MOTA changes by at most 0.21 percentage points. Since HIVE–ECIA preserves the original detections and bounding boxes, these results indicate that its gains arise primarily from improved identity continuity rather than detection or localization.
| Tracker | Identity Recovery | HOTA↑ | IDF1↑ | AssA↑ | MOTA↑ | IDSW↓ |
| QuadTrack (held-out, 512 × 120) | ||||||
| SORT [3] | Raw tracker | 26.78 | 32.20 | 26.57 | 31.40 | 135 |
| HIVE–ECIA (Ours)† | 28.79 (+2.01) | 36.01 (+3.81) | 30.68 (+4.11) | 31.54 (+0.14) | 113 (-22) | |
| ByteTrack [27] | Raw tracker | 30.55 | 36.23 | 31.93 | 34.65 | 101 |
| HIVE–ECIA (Ours)† | 30.89 (+0.34) | 36.78 (+0.55) | 32.64 (+0.71) | 34.78 (+0.13) | 79 (-22) | |
| OC-SORT [6] | Raw tracker | 26.99 | 32.99 | 28.15 | 30.23 | 95 |
| HIVE–ECIA (Ours)† | 27.67 (+0.68) | 34.65 (+1.66) | 29.56 (+1.41) | 30.34 (+0.11) | 78 (-17) | |
| BoT-SORT [1] | Raw tracker | 36.18 | 44.76 | 43.62 | 35.38 | 92 |
| HIVE–ECIA (Ours) | 37.25 (+1.07) | 46.63 (+1.87) | 46.28 (+2.66) | 35.47 (+0.09) | 78 (-14) | |
| DualPano-MOT (zero-shot, 640 × 320) | ||||||
| SORT [3] | Raw tracker | 33.58 | 35.08 | 22.25 | 55.34 | 103 |
| HIVE–ECIA (Ours)† | 40.27 (+6.69) | 41.82 (+6.74) | 31.89 (+9.64) | 55.52 (+0.18) | 88 (-15) | |
| ByteTrack [27] | Raw tracker | 34.04 | 32.25 | 21.88 | 56.74 | 85 |
| HIVE–ECIA (Ours)† | 35.56 (+1.52) | 35.44 (+3.19) | 23.85 (+1.97) | 56.84 (+0.10) | 76 (-9) | |
| OC-SORT [6] | Raw tracker | 38.79 | 41.20 | 30.60 | 54.47 | 77 |
| HIVE–ECIA (Ours)† | 40.11 (+1.32) | 43.82 (+2.62) | 32.70 (+2.10) | 54.68 (+0.21) | 59 (-18) | |
| BoT-SORT [1] | Raw tracker | 34.22 | 35.20 | 22.73 | 56.86 | 94 |
| HIVE–ECIA (Ours) | 35.25 (+1.03) | 37.22 (+2.02) | 24.08 (+1.35) | 56.91 (+0.05) | 89 (-5) |
Table 2. Comparison with MOT baselines on QuadTrack and DualPano-MOT. Within each dataset, baseline–recovery pairs share identical YOLO11l detections and tracking geometry. HIVE–ECIA modifies only canonical identity assignments. † indicates tracker-family transfer without retraining. Green values indicate improvements over the corresponding raw tracker. HOTA, IDF1, AssA, and MOTA are reported in percentages; IDSW denotes identity-switch counts. Gray rows highlight our primary configuration.
Beyond the primary BoT-SORT configuration, we evaluate transferability across 24 detector–tracker–dataset configurations (three detectors, four trackers, and two datasets) without retraining HIVE or ECIA, as detailed in Supplementary Sec. 8.3. On zero-shot DualPano-MOT, SORT gains 6.74 points in IDF1 and 9.64 points in AssA, while OC-SORT reduces IDSW by 23.4%. The varying gains across trackers suggest that recovery effectiveness depends on the association errors and available recovery candidates of the underlying tracker. Thus, HIVE–ECIA provides complementary identity correction rather than replacing conventional tracking, motivating the component analysis in § 5.3.
Ablation Studies
We evaluate the contributions of HIVE–ECIA using four configurations: A0, LR-only tracking; A1, ECIA Base recovery; A2, rule-triggered HIVE and ECIA Full; and A3, the complete Base-Guided HIVE–ECIA framework.
Component contributions and WHEN. Table 3 shows that Base-Guided querying reduces HR requests from 120 to 29 on QuadTrack (75.8%) and from 53 to 18 on
| Dataset | Variant | HOTA↑ | IDF1↑ | IDSW↓ | HR↓ |
| QuadTrack | A0 | 36.18 | 44.76 | 92 | 0 |
| QuadTrack | A1 | 36.32 | 44.85 | 86 | 0 |
| QuadTrack | A2 | 37.22 | 46.56 | 82 | 120 |
| QuadTrack | A3 (Ours) | 37.25 | 46.63 | 78 | 29 |
| DualPano | A0 | 34.22 | 35.20 | 94 | 0 |
| DualPano | A1 | 34.68 | 35.20 | 91 | 0 |
| DualPano | A2 | 35.62 | 36.96 | 85 | 53 |
| DualPano | A3 (Ours) | 35.25 | 37.22 | 89 | 18 |
Table 3. Component ablation. A0: LR-only; A1: ECIA Base; A2: rule-triggered HIVE + ECIA Full; A3: Base-Guided HIVE–ECIA (ours). HR denotes the number of HR requests.
DualPano-MOT (66.0%), compared with rule-triggered recovery. It slightly improves HOTA, IDF1, and IDSW on QuadTrack. On DualPano-MOT, IDF1 increases by 0.26 points, while HOTA and IDSW show modest regressions. These results demonstrate substantial HR acquisition savings with dataset-dependent tracking trade-offs.
HR fusion (HOW). Under matched recovery opportunities, HR-conditioned fusion improves IDF1 over Base-only on both datasets. Learned gating does not consistently outperform ungated HR Gain (): it achieves higher IDF1 on DualPano-MOT but slightly underperforms on QuadTrack. Detailed matched-ablation results are reported in the supplementary material.
Historical selection (WHICH). Table 4 demonstrates that historical evidence quality directly affects downstream identity recovery.
| Dataset | Selection | IDF1↑ | IDSW↓ |
| QuadTrack | Nearest | 44.92 | 84 |
| QuadTrack | LR heuristic | 46.26 | 81 |
| QuadTrack | HIVE (Ours) | 46.63 | 78 |
| DualPano | Nearest | 36.60 | 91 |
| DualPano | LR heuristic | 37.22 | 90 |
| DualPano | HIVE (Ours) | 37.22 | 89 |
Table 4. Historical evidence selection ablation (WHICH). Nearest: most recent historical candidate; LR heuristic: original candidate ordering; HIVE: learned historical ranking (ours). All variants use matched query opportunities and the same ECIA Full checkpoint.
identity recovery. HIVE consistently outperforms nearest-frame selection and improves association over heuristic ranking on QuadTrack. On DualPano-MOT, HIVE matches heuristic IDF1 while reducing identity switches from 90 to 89. These results support learned historical selection over purely temporal retrieval, with dataset-dependent gains relative to heuristic ranking.
Historical candidate ranking (WHICH). On 92 ground-truth-labeled identity-switch events from held-out QuadTrack sequences, HIVE achieves 63.04% top-1 selection accuracy, outperforming nearest-frame selection (27.17%) and the original LR candidate ordering (52.17%). This indicates that temporal proximity alone is insufficient for identifying useful historical evidence. However, candidate-pool oracle coverage is 86.96%, revealing an upper bound imposed by candidate availability. These results evaluate evidence selection independently of downstream identity recovery.
Bandwidth–Accuracy Trade-off
We evaluate selective HR acquisition under an emulated edge–server setting, where HR panoramas are buffered at the edge and only LR frames are continuously transmitted. HIVE–ECIA requests historical HR person crops on demand. We compare against LR-only, continuous Full-HR, and periodic-HR acquisition using BoT-SORT. Periodic-HR transmits one HR panorama every five frames, replacing the corresponding LR frame. Transmission cost is estimated using JPEG-encoded image payloads (quality 90).
Acquisition efficiency. As shown in Table 5, HIVE–ECIA reduces encoded data volume by 89.22% on QuadTrack and 79.29% on DualPano-MOT relative to Full-HR. More importantly, compared with LR-only, it improves IDF1 by 1.87 and 2.02 points, respectively, while requiring only 0.16 and 0.05 MB of additional HR data. Against periodic-HR, HIVE–ECIA achieves higher IDF1 with 62.3% less data on QuadTrack. On DualPano-MOT, periodic-HR provides higher IDF1 but transmits 76.7% more data than HIVE–ECIA. These results highlight the dataset-dependent accuracy–bandwidth trade-off of selective historical retrieval.
| Acquisition | IDF1↑ | IDSW↓ | MB↓ | Saving↑ | FPS↑ | Speedup↑ |
| QuadTrack | ||||||
| LR-only | 44.76 | 92 | 51.31 | 89.26% | 117.26 ± 13.81 | 1.44 ± 0.13× |
| Full-HR | 48.30 | 376 | 477.64 | 0.00% | 81.64 ± 4.19 | 1.00× |
| Periodic-HR () | 43.00 | 198 | 136.63 | 71.39% | 94.49 ± 0.55 | 1.16 ± 0.06× |
| HIVE–ECIA (Ours) | 46.63 | 78 | 51.47 | 89.22% | 104.01 ± 2.48 | 1.28 ± 0.06× |
| DualPano-MOT | ||||||
| LR-only | 35.20 | 94 | 181.11 | 79.29% | 112.20 ± 18.38 | 2.23 ± 0.29× |
| Full-HR | 30.38 | 182 | 874.61 | 0.00% | 50.22 ± 1.93 | 1.00× |
| Periodic-HR () | 39.62 | 88 | 320.08 | 63.40% | 79.85 ± 10.14 | 1.59 ± 0.15× |
| HIVE–ECIA (Ours) | 37.22 | 89 | 181.16 | 79.29% | 106.05 ± 18.29 | 2.11 ± 0.30× |
Table 5. Bandwidth–accuracy and runtime comparison with BoT-SORT. Data volume is measured using JPEG quality 90; savings are relative to continuous Full-HR. FPS and paired speedup are reported as mean ± sample standard deviation over three full-dataset runs on an RTX 5090.
Full-HR behavior. Unlike HIVE–ECIA, continuous Full-HR changes detector outputs and tracking associations at every frame. It produces more detections and track hypotheses, which can increase trajectory fragmentation under fixed association thresholds. This is consistent with the observed increase in IDSW despite improved detection coverage.
Computational efficiency. HIVE–ECIA achieves 104.01 ± 2.48 and 106.05 ± 18.29 FPS on QuadTrack and DualPano-MOT, respectively, corresponding to paired speedups of 1.28 ± 0.06× and 2.11 ± 0.30× over Full-HR. This reflects the benefit of avoiding continuous full-frame HR inference, although selective recovery introduces additional computation over LR-only.
Reported data savings are offline JPEG payload estimates, excluding control traffic and local buffering costs. Runtime measurements include detection and tracking but exclude image I/O, encoding, and network latency; they therefore characterize the evaluated processing pipelines rather than end-to-end network deployment.
Conclusion
We presented HIVE–ECIA, an identity-aware framework for bandwidth-efficient panoramic multi-object tracking. By jointly addressing when to acquire high-resolution evidence, which historical observation to retrieve, and how to incorporate it into identity association, our approach enables selective identity recovery without altering the underlying tracking geometry. We further introduced DualPano-MOT, a real-world dual-fisheye panoramic dataset designed to evaluate identity continuity under challenging pedestrian interactions. Experiments across four tracking methods demonstrate consistent improvements in aggregate identity association metrics, reducing identity switches by up to 21.8% on QuadTrack and 23.4% on DualPano-MOT. Compared with continuous full-resolution transmission, HIVE–ECIA reduces JPEG-encoded data volume by 89.2% and 79.3%, respectively.
Limitations. Our bandwidth evaluation relies on offline JPEG-based acquisition simulation rather than measured network transmission. Moreover, historical evidence recovery depends on the availability of informative candidate observations, while the current DualPano-MOT collection covers a limited set of controlled scenarios. Future work will investigate deployment under real network conditions, more diverse panoramic environments, and improved evidence selection under severe identity ambiguity.
LLM Usage Statement
OpenAI ChatGPT and Codex were used to assist with manuscript organization, language editing, LaTeX formatting, architecture figure refinement, code inspection, and experimental audit support. All mathematical formulations, experimental results, and scientific claims were reviewed and verified by the authors, who retain full responsibility for the submitted work.
DualPano-MOT Data Acquisition and Annotation
Data acquisition. DualPano-MOT is collected using a stationary dual-fisheye camera setup designed to capture a panoramic field of view. Two oppositely oriented fisheye cameras record complementary views of the same scene, as illustrated in Figure 4. The captured frames are temporally aligned, geometrically remapped, and stitched into 360° equirectangular panoramas. The resulting panoramic frames retain the correspondence to their original fisheye observations and serve as the high-resolution (HR) reference images. Low-resolution (LR) inputs are generated through spatial downsampling of the corresponding panoramas.

Figure 4. Dual-fisheye data acquisition. Two complementary fisheye views used to construct DualPano-MOT. The views are geometrically remapped and stitched into a 360° equirectangular panorama.
Annotation protocol. We adopt a model-assisted annotation pipeline to construct the pedestrian tracking ground truth. Initial bounding boxes and track identities are generated automatically and subsequently imported into the Computer Vision Annotation Tool (CVAT). All annotations are manually inspected and corrected, including inaccurate bounding boxes, missed detections, and inconsistent identity assignments. Figure 5 illustrates a representative annotated equirectangular panorama with pedestrian bounding boxes and identity labels.

Figure 5. DualPano-MOT annotation example. A stitched equirectangular panoramic frame with pedestrian bounding boxes and identity labels. Different colors denote distinct identities. Initial annotations are generated using a model and subsequently corrected and verified manually in CVAT to ensure consistent identities throughout each sequence.
Particular attention is paid to pedestrian crossings, partial and complete occlusions, and transitions across the panoramic boundary. Identity labels are corrected to maintain consistency throughout each sequence, including when pedestrians temporarily disappear and reappear. Frames in which a pedestrian is fully invisible are not assigned artificial bounding boxes.
Ethical considerations and consent. DualPano-MOT was collected in a controlled environment for research purposes. All participants were informed about the recording process and voluntarily consented to participate in the study. No formal institutional ethics review was sought, as the recording was conducted as a limited technical evaluation of pedestrian tracking rather than an investigation of participants’ personal characteristics. The collected videos and annotations are used to evaluate panoramic multi-object tracking and identity consistency.
Dataset composition. The collection covers five identity-oriented scenarios: normal walking, pedestrian crossing, panorama-seam interaction, partial occlusion, and full occlusion. These scenarios are designed to capture common sources of identity ambiguity in panoramic tracking.
DualPano-MOT contains 6,969 stitched frames, of which 4,969 frames are annotated, comprising 8,518 pedestrian bounding boxes and 10 sequence-local identity trajectories. The annotations also contain 74 within-track observation gaps, including nine gaps of at least 30 frames. Such gaps are particularly relevant to identity recovery, as they represent intervals in which a target is not observed before subsequently reappearing.
Unlike QuadTrack, which emphasizes panoramic tracking under moving-camera conditions, DualPano-MOT focuses on identity continuity under controlled pedestrian interactions using a stationary dual-fisheye setup. The dataset therefore provides a complementary evaluation setting for historical evidence-based identity recovery.
Implementation Details and Additional Experiments
Figure 6 provides a module-level view of WHEN (Eqs. (2)–(7)), WHICH (Eqs. (8) and (10)}) and HOW (Eqs. (11)–(14)` and (16)}).

Figure 6. Detailed architecture of HIVE–ECIA. WHEN: ECIA Base-constructed HR querying. WHICH: HIVE ranks historical LR candidates using query context, candidate-specific appearance features, and temporal metadata. HOW: ECIA Full integrates the retrieved HR evidence through residual gain and learned gated fusion. Accepted corrections update canonical identities while preserving tracking geometry.
Training and evaluation protocol. We use sequence-disjoint QuadTrack partitions for model development. The training set contains sequences 0015, 0018, 0020–0024, 0026–0027, and 0029, while sequences 0016, 0017, 0025, and 0031 are reserved for held-out evaluation. HIVE uses sequences 0019, 0028, and 0030 for validation. Recovery-aware ECIA uses 0019 and 0028 for model validation, with 0030 designated for decision calibration. ECIA Base is trained first and subsequently frozen during ECIA Full optimization.
All five DualPano-MOT sequences are evaluated zero-shot, without dataset-specific model updates. The appearance encoder, HIVE and ECIA checkpoints remain frozen throughout evaluation. Ground truth is used for offline supervision and metric computation, but is unavailable during online identity recovery.
Decision calibration. The final ECIA merge probability and margin thresholds (0.80 and 0.40) were selected from a 49-configuration validation sweep on QuadTrack sequence 0030. In contrast, the Base-Guided HR query thresholds were fixed heuristic settings, without a separate explicit validation sweep. All settings were held fixed for held-out QuadTrack and zero-shot DualPano-MOT evaluation. Sequence 0030 was also used during HIVE model validation; therefore, the final merge calibration is not fully independent of HIVE checkpoint selection.
Base-Guided query settings. For Eq. (6), the fixed query-policy thresholds are . These values were fixed before held-out evaluation and were not optimized through a separate query-policy validation sweep.
Implementation settings. Table 6 summarizes the training configurations. HIVE selects historical observations from at most 20 LR candidates. ECIA considers at most 12 competing identities and retrieves no more than one HR person crop per query.
| Parameter | HIVE | ECIA Full |
| Random seed | 42 | 42 |
| Max. epochs | 100 | 50 |
| Batch size | 8 | 128 |
| Learning rate | ||
| Weight decay | ||
| Gradient clipping | 5.0 | 1.0 |
| Early-stop patience | 10 | 10 |
Table 6. Training configurations. Values are specified for HIVE v2 and recovery-aware ECIA Full.
HIVE training supervision. HIVE is trained on GT-confirmed identity-switch events from QuadTrack. For each event at frame , an LR-only historical candidate pool is constructed from frames . At most 20 candidates are selected using appearance-hard, recent, and
temporally diverse proposals, without using GT to rank candidate observations. After candidate selection, Hungarian matching at IoU assigns positive labels to observations matching the event’s GT identity; other-identity and unmatched observations are labeled negative.
We optimize Eq. (9) with and a positive-class BCE weighting of 2.253. The ranking loss sees all within-event positive–negative pairs and excludes events lacking either class. GT is used only for offline supervision; online event triggering and HIVE ranking use causal LR observations without GT or HR features. HR retrieval occurs only after the historical candidate is selected.
Recovery-Aware ECIA Training Objective
We detail the recovery-aware objective introduced in Eq. (15). Following the fusion formulation in Eq. (13), the Base projection, association head, and anticipation head remain frozen. Only HR Gain and Evidence Gate are opti mized.
Training targets. For each offline training event , let denote the valid decision classes: the predecessor, competitors, and abstention. Let be the ground-truth decision class.
The complete Base and fused decision logits, including abstention, are and . Their softmax probabilities are and , respectively. Invalid padded competitor classes are masked.
Let denote the offline sample type, indicate HR availability, and denote the sample weight. We define the Base-correct and recovery-eligible indicators as
A positive sample has the predecessor as its correct class and verified HR evidence supporting that identity. Thus, identifies an HR-available event where the frozen Base is incorrect and HR evidence supports recovery.
Decision objectives. We define the per-event losses as
where , , and BCE denotes binary cross-entropy. The wrong-merge loss directly supervises whether the target should be merged with the predecessor.
Masked supervision. For per-event losses , a binary mask , and sample weights , define
where . An empty eligible set contributes zero. The corresponding objectives are
Preservation penalizes reductions in correct-class probability when Base already predicts correctly. Recovery emphasizes classification errors on verified HR-supported events missed by Base.
Gate supervision. Let denote the learned Evidence Gate and its binary supervision target. With , the gate objective is
Only HR-available events contribute. Recovery-eligible events receive a higher relative weight than events where HR influence should be suppressed.
Optimization weights. The coefficients in Eq. (15) are . Sample weights are 1.0, 1.0, 2.0, and 1.5 for positive, negative, hard-negative, and no-merge events, respectively. Ground-truth-derived targets and eligibility masks are used only during offline training.
Matched HR Fusion Ablation
We isolate the contribution of HR Gain and Evidence Gate under matched recovery opportunities. We compare three variants: Base-only (), ungated HR Gain (), and learned Evidence Gate.
All variants share identical raw tracking outputs, candidate events, and evaluation sequences. The two HR-enabled variants use the same frozen checkpoint, historical timestamps, and retrieved HR crops. No variant-specific retraining or threshold tuning is used.
Discussion. HR-conditioned fusion improves IDF1 over Base-only on both datasets. On QuadTrack, ungated fusion slightly outperforms learned gating across all reported metrics. On DualPano-MOT, the learned gate achieves higher IDF1 (37.22 vs. 36.61), whereas ungated fusion yields higher HOTA and AssA and fewer identity switches. These results confirm the utility of historical HR evidence while revealing dataset- and metric-dependent trade-offs in learned evidence modulation.
Detector–Tracker Generalization
To evaluate transferability beyond the primary YOLO11-L configuration, we additionally consider YOLOv10-L and
| Dataset | Variant | HOTA↑ | IDF1↑ | AssA↑ | IDSW↓ |
| QuadTrack | Base-only | 36.32 | 44.85 | 43.96 | 86 |
| QuadTrack | Ungated () | 37.29 | 46.67 | 46.37 | 75 |
| QuadTrack | Learned gate | 37.25 | 46.63 | 46.28 | 78 |
| DualPano | Base-only | 34.68 | 35.20 | 23.32 | 91 |
| DualPano | Ungated () | 35.36 | 36.61 | 24.23 | 87 |
| DualPano | Learned gate | 35.25 | 37.22 | 24.08 | 89 |
Table 7. Matched HR fusion ablation with YOLO11-L and BoT-SORT. HOTA, IDF1, and AssA are percentages. All variants use matched recovery opportunities. HR crops are available for 28/29 queries on QuadTrack and 18/18 on DualPano-MOT.
YOLO12-L with SORT, ByteTrack, OC-SORT, and BoT-SORT. For each detector–dataset pair, all four trackers consume the same immutable detection cache. Each recovery branch shares its baseline’s detections and raw tracking geometry.
The frozen HIVE, ECIA, OSNet-AIN, Base-Guided query policy, and safety mechanisms remain unchanged. Person detection uses confidence 0.20 and NMS IoU 0.70. Physical LR resolutions are 512 120 for QuadTrack and 640 320 for DualPano-MOT. All results are evaluated at IoU 0.50 without ground-truth access during inference. HR payload denotes the actual JPEG-Q90 encoded crops retrieved by HIVE–ECIA.
Overall transferability. Table 8 and Table 9 demonstrate consistent improvements across 24 detector–tracker–dataset configurations. HIVE–ECIA improves HOTA and AssA while reducing IDSW in all 24 comparisons. IDF1 improves in 22 configurations, including 11 of 12 on each dataset. The two exceptions are YOLOv10-L with SORT on QuadTrack, where IDF1 decreases by 0.33 points, and YOLO12-L with BoT-SORT on DualPano-MOT, where IDF1 remains unchanged at the reported precision.
| Detector | Tracker | HOTA↑ | IDF1↑ | AssA↑ | IDSW↓ |
| YOLOv10-L | SORT | 24.79 → 24.84 +0.05 | 30.69 → 30.36 −0.33 | 29.76 → 29.87 +0.12 | 81 → 71 −10 |
| YOLOv10-L | ByteTrack | 27.45 → 28.39 +0.94 | 31.85 → 33.29 +1.44 | 34.05 → 36.43 +2.38 | 84 → 63 −21 |
| YOLOv10-L | OC-SORT | 25.66 → 25.84 +0.18 | 30.47 → 30.62 +0.16 | 35.07 → 35.57 +0.50 | 65 → 58 −7 |
| YOLOv10-L | BoT-SORT | 30.24 → 30.32 +0.08 | 34.56 → 35.13 +0.57 | 39.97 → 40.22 +0.25 | 91 → 69 −22 |
| YOLO11-L | SORT | 26.78 → 28.79 +2.01 | 32.20 → 36.01 +3.81 | 26.57 → 30.68 +4.11 | 135 → 113 −22 |
| YOLO11-L | ByteTrack | 30.55 → 30.89 +0.34 | 36.23 → 36.78 +0.55 | 31.93 → 32.64 +0.71 | 101 → 79 −22 |
| YOLO11-L | OC-SORT | 26.99 → 27.67 +0.68 | 32.99 → 34.65 +1.66 | 28.15 → 29.56 +1.41 | 95 → 78 −17 |
| YOLO11-L | BoT-SORT | 36.18 → 37.25 +1.07 | 44.76 → 46.63 +1.87 | 43.62 → 46.28 +2.66 | 92 → 78 −14 |
| YOLO12-L | SORT | 20.28 → 22.11 +1.83 | 23.68 → 27.17 +3.49 | 21.37 → 25.39 +4.02 | 88 → 76 −12 |
| YOLO12-L | ByteTrack | 22.31 → 23.80 +1.49 | 26.00 → 28.94 +2.94 | 23.49 → 26.70 +3.21 | 96 → 84 −12 |
| YOLO12-L | OC-SORT | 24.39 → 24.47 +0.08 | 31.08 → 31.28 +0.20 | 34.02 → 34.23 +0.21 | 50 → 46 −4 |
| YOLO12-L | BoT-SORT | 27.78 → 28.68 +0.90 | 32.51 → 34.84 +2.33 | 35.48 → 37.73 +2.25 | 109 → 92 −17 |
Table 8. Detector–tracker generalization on held-out QuadTrack. Each cell reports Raw → HIVE–ECIA, with the change shown in green. HOTA, IDF1, and AssA are reported in percentages; IDSW denotes switch counts. The gray row marks the primary configuration used in the main paper.
| Detector | Tracker | HOTA↑ | IDF1↑ | AssA↑ | IDSW↓ |
| YOLOv10-L | SORT | 36.87 → 41.13 +4.26 | 39.87 → 44.72 +4.84 | 24.96 → 30.99 +6.03 | 86 → 71 −15 |
| YOLOv10-L | ByteTrack | 38.98 → 42.98 +4.00 | 40.09 → 44.50 +4.41 | 26.80 → 32.54 +5.74 | 79 → 73 −6 |
| YOLOv10-L | OC-SORT | 36.79 → 43.17 +6.38 | 40.11 → 47.42 +7.31 | 25.57 → 35.11 +9.54 | 78 → 61 −17 |
| YOLOv10-L | BoT-SORT | 37.32 → 40.89 +3.57 | 39.55 → 44.37 +4.81 | 25.45 → 30.48 +5.03 | 101 → 94 −5 |
| YOLO11-L | SORT | 33.58 → 40.27 +6.69 | 35.08 → 41.82 +6.74 | 22.25 → 31.89 +9.64 | 103 → 88 −15 |
| YOLO11-L | ByteTrack | 34.04 → 35.56 +1.52 | 32.25 → 35.44 +3.19 | 21.88 → 23.85 +1.97 | 85 → 76 −9 |
| YOLO11-L | OC-SORT | 38.79 → 40.11 +1.32 | 41.20 → 43.82 +2.62 | 30.60 → 32.70 +2.10 | 77 → 59 −18 |
| YOLO11-L | BoT-SORT | 34.22 → 35.25 +1.03 | 35.20 → 37.22 +2.02 | 22.73 → 24.08 +1.35 | 94 → 89 −5 |
| YOLO12-L | SORT | 37.40 → 42.34 +4.94 | 39.06 → 44.18 +5.12 | 27.00 → 34.51 +7.51 | 67 → 55 −12 |
| YOLO12-L | ByteTrack | 36.49 → 37.15 +0.66 | 36.58 → 37.45 +0.87 | 24.74 → 25.63 +0.89 | 73 → 69 −4 |
| YOLO12-L | OC-SORT | 36.05 → 41.34 +5.29 | 38.00 → 44.48 +6.47 | 25.89 → 33.95 +8.06 | 49 → 45 −4 |
| YOLO12-L | BoT-SORT | 36.02 → 36.04 +0.02 | 37.15 → 37.15 − | 24.95 → 24.97 +0.02 | 106 → 101 −5 |
Table 9. Detector–tracker generalization on DualPano-MOT (zero-shot). Each cell reports Raw → HIVE–ECIA, with the change shown in green. HOTA, IDF1, and AssA are reported in percentages; IDSW denotes switch counts. The gray row marks the primary configuration used in the main paper.
Detector-dependent gains. The improvements vary with the underlying detector and tracker. On zero-shot DualPano-MOT, YOLOv10-L with OC-SORT achieves a 7.31-point IDF1 improvement and reduces IDSW from 78 to 61. Conversely, YOLO12-L with BoT-SORT shows marginal changes in association metrics despite fewer identity switches. These differences reflect the dependence of identity recovery on the available association candidates and ambiguities. Although the acquisition policy remains fixed, the resulting recovery opportunities may differ across configurations.
Overall, the results support detector–tracker transferability without retraining HIVE or ECIA. Within each paired comparison, detections and tracking geometry remain unchanged, isolating the effect of canonical identity reassociation.
Qualitative Results
We provide qualitative comparisons of HIVE–ECIA against four conventional MOT trackers, namely BoT-SORT, ByteTrack, OC-SORT, and SORT, on QuadTrack and our DualPano-MOT dataset. The examples focus on identity consistency across temporally separated observations, where conventional trackers may assign new identities or lose the target track. In the visualizations, green indicates identity preservation by our method, yellow highlights identity switches, and red denotes tracking failures. Identity labels are compared across frames within each method, rather than across different trackers. QuadTrack. Figure 7 presents a representative identity recovery case in an outdoor panoramic sequence. Between frames 297 and 301, HIVE–ECIA consistently assigns ID 50 to the target pedestrian. In comparison, BoT-SORT changes the target identity from 50 to 102, ByteTrack from 93 to 170, and SORT from 73 to 19. OC-SORT fails to maintain the target track in the later observation. These results illustrate how identity fragmentation can occur even when the target remains visually observable, and how our recovery framework can preserve a consistent identity.

Figure 7. Qualitative comparison on QuadTrack. Tracking results at frames 297 and 301 are shown for HIVE–ECIA and four conventional MOT trackers. Our method maintains the target identity (ID 50) across the two observations. In contrast, BoT-SORT, ByteTrack, and SORT assign different identities to the same target, while OC-SORT fails to maintain the corresponding track in the later frame. Green indicates preserved identity, yellow highlights identity switches, and red denotes track loss.
DualPano-MOT. Figure 8 further evaluates identity consistency in our stationary panoramic setting. Between frames 895 and 903, HIVE–ECIA maintains the same target identity (ID 29), whereas BoT-SORT changes its identity from 29 to 36 and ByteTrack from 57 to 73. Meanwhile, OC-SORT and SORT fail to maintain the corresponding target track in the later frame. This example highlights the challenges of maintaining identity continuity in panoramic observations and the benefit of incorporating historical appearance evidence during recovery.

Figure 8. Qualitative comparison on DualPano-MOT. Tracking results at frames 895 and 903 illustrate identity consistency in an indoor panoramic sequence. HIVE–ECIA preserves the target identity (ID 29) across the two observations. BoT-SORT and ByteTrack exhibit identity switches, whereas OC-SORT and SORT lose the corresponding target track in the later frame. Green, yellow, and red indicate identity preservation, identity switches, and track loss, respectively.
Discussion. Across both examples, HIVE–ECIA preserves the target identity where the conventional trackers exhibit identity switches or track loss. Unlike approaches that rely solely on current-frame observations, our framework can selectively acquire historical high-resolution appearance evidence to resolve ambiguous associations. Importantly, identity recovery is performed without modifying the underlying tracker’s detection boxes or geometric trajectories. These visualizations illustrate representative recovery outcomes; the quantitative improvements across complete sequences are reported in the main paper.
References
- [1]Nir Aharon, Roy Orfaig, and Ben-Zion Bobrovsky. Bot-sort: Robust association multi-pedestrian tracking, 2022. 2, 7arxiv.org/abs/2206.14651
- [2]Keni Bernardin and Rainer Stiefelhagen. Evaluating multiple object tracking performance: The clear mot metrics. EURASIP Journal on Image and Video Processing, 2008:246309, 2008. 6
- [3]Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE International Conference on Image Processing (ICIP), pages 3464–3468, 2016. 2, 7DOI
- [4]Christopher J. C. Burges, Tal Shaked, Erin Renshaw, Ari Lazier, Matt Deeds, Nicole Hamilton, and Gregory N. Hullen-der. Learning to rank using gradient descent. In Proceedings of the International Conference on Machine Learning, pages 89–96, 2005. 4DOI
- [5]Jiarui Cai, Mingze Xu, Wei Li, Yuanjun Xiong, Wei Xia, Zhouwen Tu, and Stefano Soatto. Memot: Multi-object tracking with memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8090–8100, 2022. 2arxiv.org/abs/2203.16761
- [6]Jinkun Cao, Jiangmiao Pang, Xinshuo Weng, Rawal Khi-rodkar, and Kris Kitani. Observation-centric sort: Rethink-ing sort for robust multi-object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9686–9696, 2023. 2, 7arxiv.org/abs/2203.14360
- [7]Patrick Dendorfer, Aljosa Osep, Anton Milan, Konrad Schindler, Daniel Cremers, Ian Reid, Stefan Roth, and Laura Leal-Taixe. Motchallenge: A benchmark for single-camera multiple target tracking. International Journal of Computer Vision, 129(4):845–881, 2021. 6arxiv.org/abs/2010.07548
- [8]Kuntai Du, Ahsan Pervaz, Xin Yuan, Aakanksha Chowdhery, Qizhen Zhang, Henry Hoffmann, and Junchen Jiang. Server-driven video streaming for deep learning inference. In Proceedings of the ACM Special Interest Group on Data Communication (SIGCOMM), 2020. 2, 3DOI
- [9]Ruopeng Gao and Limin Wang. Memotr: Long-term memory-augmented transformer for multi-object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9901–9910, 2023. 2arxiv.org/abs/2307.15700
- [10]Ruopeng Gao, Ji Qi, and Limin Wang. Multiple object tracking as id prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27883–27893, 2025. 2arxiv.org/abs/2403.16848
- [11]Huajian Huang, Yinzhe Xu, Yingshu Chen, and Sai-Kit Ye-ung. 360VOT: A new benchmark dataset for omnidirectional visual object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20566–20576, 2023. 2, 6arxiv.org/abs/2307.14630
- [12]Junchen Jiang, Ganesh Ananthanarayanan, Peter Bodik, Siddhartha Sen, and Ion Stoica. Chameleon: Scalable adapta-tion of video analytics. In Proceedings of ACM SIGCOMM, pages 253–266, 2018. 3DOI
- [13]Yuanqi Li, Arthi Padmanabhan, Pengzhan Zhao, Yufei Wang, Guoqing Harry Xu, and Ravi Netravali. Reducto: On-camera filtering for resource-efficient real-time video analyt-ics. In Proceedings of the ACM SIGCOMM 2020 Conference, pages 359–376, 2020. 3DOI
- [14]Jianhua Lin. Divergence measures based on the shannon entropy. IEEE Transactions on Information Theory, 37(1):145–151, 1991. 5DOI
- [15]Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. International Journal of Computer Vision, 129(2):548–578, 2021. 6DOI
- [16]Kai Luo, Hao Shi, Kunyu Peng, Fei Teng, Sheng Wu, Kaiwei Wang, and Kailun Yang. Omnitrack++: Omnidirectional multi-object tracking by learning large-fov trajectory feed-back, 2025. 2arxiv.org/abs/2511.00510
- [17]Kai Luo, Hao Shi, Sheng Wu, Fei Teng, Mengfei Duan, Chang Huang, Yuhang Wang, Kaiwei Wang, and Kailun Yang. Omnidirectional multi-object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21959–21969, 2025. 1, 2, 6arxiv.org/abs/2503.04565
- [18]Gerard Maggiolino, Adnan Ahmad, Jinkun Cao, and Kris Kitani. Deep oc-sort: Multi-pedestrian tracking by adaptive re-identification. In IEEE International Conference on Image Processing, pages 3025–3029, 2023. 2
- [19]Roberto Martin-Martin, Mihir Patel, Hamid Rezatofighi, Abhijeet Shenoi, JunYoung Gwak, Eric Frankel, Amir Sadighian, and Silvio Savarese. JRDB: A dataset and benchmark of egocentric robot visual perception of humans in built environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):6748–6765, 2023. 2, 6
- [20]Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixé, and Christoph Feichtenhofer. TrackFormer: Multi-object tracking with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8844–8854, 2022. 2arxiv.org/abs/2101.02702
- [21]Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 164–173, 2021. 2arxiv.org/abs/2006.06664
- [22]Zheng Qin, Sanping Zhou, Le Wang, Jinghai Duan, Gang Hua, and Wei Tang. Motriack: Learning robust short-term and long-term motions for multi-object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17939–17948, 2023. 2arxiv.org/abs/2303.10404
- [23]Ergys Ristani, Francesco Solera, Roger S. Zou, Rita Cuc-chiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In Computer Vision – ECCV 2016 Workshops, pages 17–35. Springer, 2016. 6
- [24]Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In 2017 IEEE International Conference on Image Processing (ICIP), 2017. 2arxiv.org/abs/1703.07402
- [25]Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xiangyu Zhang, and Yichen Wei. MOTR: End-to-end multiple-object tracking with transformer. In Computer Vision – ECCV 2022, pages 659–675, 2022. 2arxiv.org/abs/2105.03247
- [26]Ben Zhang, Xin Jin, Sylvia Ratnasamy, John Wawrzÿnek, and Edward A. Lee. Awstream: Adaptive wide-area streaming analytics. In Proceedings of the ACM Special Interest Group on Data Communication (SIGCOMM), 2018. 2, 3DOI
- [27]Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In Computer Vision – ECCV 2022, 2022. 2, 7arxiv.org/abs/2110.06864
- [28]Yuang Zhang, Tiancai Wang, and Xiangyu Zhang. MOTRv2: Bootstrapping end-to-end multi-object tracking by pre-trained object detectors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22056–22065, 2023. 2DOI
- [29]Kaiyang Zhou, Yongxin Yang, Andrea Cavallaro, and Tao Xiang. Learning generalisable omni-scale representations for person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):5056–5069, 2022. 6DOI