betaFeedback

WorksComputer Science

Oriented Semi-DETR for Semi-Supervised Aerial Object Detection

Abstract

Semi-supervised object detection has recently been extended to DETR, where oriented boxes introduce an ambiguity that does not arise for axis-aligned boxes: the map from parameters to rectangles is two-to-one, since (cx,cy,w,h,θ)(c_x,c_y,w,h,θ) and (cx,cy,h,w,θ+π/2)(c_x,c_y,h,w,θ+π/2) denote the same rotated rectangle. Losses and matching costs are computed from the parameters, so two boxes that coincide in space can be far apart in the metric that drives training. A correctly localized student box is then penalized by the regression loss, and the inflated matching cost leads the Gaussian mixture model (GMM) filter to discard it. We propose Oriented Semi-DETR, which removes this ambiguity wherever a pseudo box enters training. Swap-aligned Target Selection regresses towards the encoding closer to the student prediction. A Swap-invariant Reliability Cost minimizes over the two encodings inside the matching cost, so the filter decides on geometry alone. Swap-aware Pseudo-label Denoising starts denoising earlier, ramps its loss weight linearly, and weights each query by the teacher score and encoding margin. All three are parameter-free and require no additional forward pass. Across all label ratios of DOTA-v2.0, Oriented Semi-DETR outperforms DETR-based supervised and semi-supervised baselines, with the largest gains at the strict IoU threshold.

main.pdf

Introduction

Aerial images contain objects at arbitrary orientations, so detectors in this domain predict oriented bounding boxes (OBBs) rather than axis-aligned boxes [1, 2]. Annotation is expensive: a single DOTA image may contain hundreds of instances, and each one needs four vertices instead of two corners [3, 4]. Semi-supervised object detection is therefore attractive in this domain, provided that pseudo-label noise does not become entangled with the box representation.

Teacher–student frameworks generate pseudo labels from a teacher view and train a student on a stronger augmentation [5, 6], and an exponential moving average keeps the teacher stable [7, 8, 6]. Detection transformers assign one query to each object [9, 10, 11]. That assignment becomes unstable once the targets are pseudo labels, so such detectors are hard to train under semi-supervision [12]. Semi-DETR [12] addresses this with stage-wise hybrid matching, cross-view query consistency, and a Gaussian mixture model (GMM) that separates reliable from unreliable pseudo boxes by their Hungarian cost. Sparse Semi-DETR [13] adds sparse learnable queries. Neither was designed for oriented boxes, and neither resolves their two equivalent encodings.

The obstacle is that the map from OBB parameters to rectangles is two-to-one. Let σ\sigma exchange ww and hh and rotate θ\theta by π/2\pi/2; then b\mathbf{b} and σ(b)\sigma(\mathbf{b}) cover the same pixels and are the two encodings of that rectangle. Supervised oriented detectors handle the resulting loss discontinuity with angular encodings or distribution-based losses [14, 15, 16, 17], and aerial-image transformers follow the same practice [18]. These methods assume a fixed annotation convention. Under semi-supervision the target is a teacher output, so the teacher and the student may adopt different encodings. A student query may reproduce the pseudo box geometrically and still receive a large penalty that the pipeline cannot distinguish from a genuine localization error. Recent semi-supervised oriented detectors weight the consistency loss by the teacher–student orientation difference [19] or combine several agreement cues [20]. Neither is built on a detection transformer, so neither has a Hungarian assignment or denoising queries, the two places our method operates on. The ambiguity matters most when pseudo labels first become useful: at 20k iterations the swapped encoding is closer for 16.9% of matched pseudo-label and query pairs. Replacing the fixed regression cost with a swap-invariant one changes 18.9% of Hungarian assignments and 7.0% of GMM filtering decisions. The effect concentrates on near-square boxes and weakens as the teacher matures, by which time it has already shaped the pseudo-label distribution.

Oriented Semi-DETR comprises three components, summarized in Figure 1. First, Swap-aligned Target Selection regresses towards whichever encoding is closer to the student prediction, at the one-to-many regression target and at the supervised and pseudo-label denoising targets. The selection is a hard min⁡\min, so the target is always a valid rectangle. Second, a Swap-invariant Reliability Cost takes the minimum over the two encodings within the matching cost, so the filter decides on geometry rather than on the encoding the teacher produced. Third, Swap-aware Pseudo-label Denoising advances the start of pseudo-label denoising, ramps its loss weight linearly, and scales each query by the teacher score qq and the encoding margin mm.

Figure 1. Overview of Oriented Semi-DETR. The teacher runs on the weakly augmented view and produces pseudo boxes, and the student runs on the strongly augmented view. Both decoders receive the same three query groups, kept apart by an attention mask: consistency queries, contrastive denoising (CDN) queries, and matching queries. An oriented box has two encodings, bb and σ(b)\sigma(b), for the same rectangle. We select the one closer to the student prediction and apply that choice at the numbered steps.

Method

Preliminaries

We build on Semi-DETR [12] with a DINO [11] detector adapted to oriented boxes. A teacher and a student share the architecture, the teacher is the exponential moving average of the student, and each iteration samples one labeled and one unlabeled image. The teacher runs on the weakly augmented view and its outputs above a score threshold become pseudo boxes. Every pseudo box is matched to a student query by a Hungarian assignment whose cost combines classification, regression and rotated IoU terms. A two-component GMM is fitted to the resulting costs and splits the pseudo boxes into a reliable and an unreliable set. Denoising (DN) queries follow DN-DETR’s perturbation of ground-truth boxes [21]; in the unlabeled branch they are built from teacher pseudo boxes, following Semi-DETR [12]. Each decoder therefore receives three query groups, separated by an attention mask: consistency, contrastive denoising, and matching queries, as shown in Figure 1. An oriented box is written b=(cx,cy,w,h,θ)\mathbf{b} = (c_x, c_y, w, h, \theta). We store the angle normalized as a=θ/π+1/2∈[0,1)a = \theta/\pi+ 1/2 \in[0, 1) and the center and size normalized by the image width and height.

The two encodings of an oriented box

Define

σ(cx,cy,w,h,θ)=(cx,cy,h,w,wrap⁡(θ+π/2)),(1)\sigma(c_x,c_y,w,h,\theta) = (c_x,c_y,h,w,\operatorname{wrap}(\theta+\pi/2)), \tag*{(1)}

where wrap returns the angle to [−π/2,π/2)[-\pi/2,\pi/2). Applying σ\sigma twice returns the original box, so every rectangle corresponds to the pair {b,σ(b)}\{b,\sigma(b)\}, whose members differ only in which side is called the width. One implementation detail matters: ww and hh are normalized by different image dimensions, so the exchange must be performed on absolute pixel values and the result normalized again. Exchanging the normalized values directly gives a different rectangle unless the image is square.

We measure the disagreement between a prediction p\mathbf{p} and a target t\mathbf{t} with

d(p,t)=∣Δcx∣+∣Δcy∣+∣Δw∣+∣Δh∣+λφ(Δa),(2)d(\mathbf{p},\mathbf{t}) = |\Delta c_x| + |\Delta c_y| + |\Delta w| + |\Delta h| + \lambda\varphi(\Delta a), \tag*{(2)}

where every term is taken on normalized coordinates, φ(Δa)=min⁡(∣Δa∣,1−∣Δa∣)∈[0,1/2]\varphi(\Delta a) = \min(|\Delta a|,1-|\Delta a|) \in[0,1/2] is the wrapped distance on the unit circle, and λ=1\lambda= 1. The same φ\varphi and λ\lambda are shared by the regression loss, the matching cost and the selection rule of § 1.2.3, so within one configuration boxes are never compared under two different metrics.

Swap-aligned target selection

Let p\mathbf{p} be a student prediction assigned to a target t\mathbf{t}. Instead of regressing towards whichever encoding the teacher produced, we regress towards

t∗=arg⁡min⁡t′∈{t,σ(t)} d(sg⁡(p),t′),(3)\mathbf{t}^{*} = \underset{\mathbf{t}' \in\{\mathbf{t},\sigma(\mathbf{t})\}}{\arg\min}\ d(\operatorname{sg}(\mathbf{p}),\mathbf{t}'), \tag*{(3)}

where sg denotes the stop-gradient and a tie is resolved in favor of the teacher’s own encoding. The selection never interpolates between the two encodings, since an interpolated box would not be a valid rectangle. Writing d0=d(p,t)d_0 = d(\mathbf{p},\mathbf{t}) and d1=d(p,σ(t))d_1 = d(\mathbf{p},\sigma(\mathbf{t})) gives d(p,t∗)=min⁡(d0,d1)≤d0d(\mathbf{p},\mathbf{t}^{*}) = \min(d_0,d_1) \le d_0, so the selected target is never farther

from the prediction than the teacher’s encoding, and the two coincide whenever the teacher’s encoding is already the closer one.

Equation (3) is applied to three regression targets: that of the queries selected by the one-to-many assigner, which carries most of the gradient, and those of the supervised and pseudo-label DN queries. On the two DN paths a perturbed copy of a box enters the decoder, which must recover the original, and the ambiguity is most harmful there, since the perturbation is small by construction and an encoding mismatch then dominates it. Keeping the two DN paths separately switchable allows us to attribute them in § 2.3.

Swap-invariant reliability cost

The GMM filter of Semi-DETR [12] decides which pseudo boxes survive, so the same ambiguity has a different consequence there: a pseudo box may be discarded because its cost is inflated by the encoding. The matching cost is a weighted sum of a classification, a regression and a rotated-IoU term, and the GMM is fitted to the total cost of the matched pairs. We replace its regression term by

Creg(p,b)=λregmin⁡(d(p,b),d(p,σ(b))).(4)C_{\mathrm{reg}}(\mathbf{p},\mathbf{b}) = \lambda_{\mathrm{reg}}\min\left(d(\mathbf{p},\mathbf{b}),d(\mathbf{p},\sigma(\mathbf{b}))\right). \tag*{(4)}

Here λreg\lambda_{\mathrm{reg}} is the regression weight of the matcher, set to 5. The penalty is substantial. The fixed matcher we compare against uses a cosine angle penalty. If the student predicts exactly σ(b)\sigma(\mathbf{b}), that term alone contributes λreg(1−cos⁡π)=10\lambda_{\mathrm{reg}}(1-\cos\pi) = 10, and the exchanged ww and hh add a further 2λreg∣w−h∣2\lambda_{\mathrm{reg}}|w-h|. A box whose normalized width and height differ by 0.15 therefore costs about 11.5 although the student has localized it perfectly, while the rotated IoU term of the same matching cost reports a perfect overlap. (4) returns 0 for the same pair. Because the GMM is refitted at every iteration, inflating one part of the cost distribution moves the decision boundary for every pseudo box, including unambiguous ones.

Swap-aware pseudo-label denoising

Denoising perturbs a target box and requires the decoder to recover it. On the labeled branch the target is an annotation, so every denoising query carries the same amount of information. On the unlabeled branch the target is a pseudo box whose reliability changes from one instance to the next. We therefore separate the activation schedule from the instance-wise weighting. Let qj∈[0,1]q_j \in[0,1] be the teacher score of the pseudo box behind the jj-th denoising query. We use a linear ramp

s(t)=min⁡(max⁡(t−t0T,0),1),(5)s(t)=\min\left(\max\left(\frac{t-t_0}{T},0\right),1\right), \tag*{(5)}

where tt is the current iteration, t0t_0 is the activation step and TT is the ramp length. The pipeline we build on enables pseudo-label denoising at 160k, at the second decay of the learning rate. We instead set t0=120kt_0=120\mathrm{k} and T=10kT=10\mathrm{k}, so denoising begins at the first decay and its weight rises over 10k iterations. We refer to this schedule as the early ramp. The first setting applies the ramp alone, so every denoising query of the iteration carries wj=s(t)w_j=s(t). The second scales the queries by the teacher score,

wxywh,j=qjs(t),wθ,jscore=qjs(t).(6)w_{xywh,j}=q_js(t),\qquad w^{\mathrm{score}}_{\theta,j}=q_js(t). \tag*{(6)}

Oriented Semi-DETR adds an encoding margin to the angle regression. For the current student DN prediction pj\mathbf{p}_j and pseudo box bj\mathbf{b}_j, we define

mj=∣d(pj,bj)−d(pj,σ(bj))∣d(pj,bj)+d(pj,σ(bj))+ε,wθ,jmargin=qjs(t)mj.(7)m_j=\frac{\left|d(\mathbf{p}_j,\mathbf{b}_j)-d(\mathbf{p}_j,\sigma(\mathbf{b}_j))\right|}{d(\mathbf{p}_j,\mathbf{b}_j)+d(\mathbf{p}_j,\sigma(\mathbf{b}_j))+\varepsilon},\qquad w^{\mathrm{margin}}_{\theta,j}=q_js(t)m_j. \tag*{(7)}

Here ε=10−6\varepsilon=10^{-6}, and mj∈[0,1)m_j\in[0,1) measures the relative separation of the two encodings under the same distance dd. It is small when the prediction is nearly indifferent between them and large when one is clearly closer. The factor qjs(t)q_js(t) weights the four spatial L1 coordinates and the per-box rotated-IoU term, while mjm_j multiplies only the angle L1 coordinate, so it changes neither the classification nor the IoU weight. All weights are computed without gradient, and the target of Equation (3) is unchanged.

Relation to prior treatments of the ambiguity

A Gaussian-based loss maps a box to a two-dimensional Gaussian, and b\mathbf{b} and σ(b)\sigma(\mathbf{b}) give the same Gaussian, so GWD [15] and KLD [16] are already invariant under σ\sigma. Substituting such a loss changes the metric everywhere, including the labeled branch, where the annotation convention fixes the encoding and the ambiguity never arises. Equation (3) leaves the metric unchanged and modifies the target alone, so both branches remain under a single loss and supervised behavior is preserved. The same arg min extends to the GMM filter, which is a decision rule rather than a loss.

RIDet [22] is the closest supervised counterpart. Its representation invariance loss also minimizes over the equivalent encodings of a box, so its target is selected during training. Two differences separate it from Equation (3). Its target is an annotation whose encoding a labeling convention already fixes, while ours is a teacher output that obeys no convention and shifts as the teacher is updated. Its minimization is also confined to the loss, so it does not reach the Hungarian cost, the reliability filter, or the denoising queries built from pseudo boxes, the three places where a semi-supervised pipeline encounters the ambiguity.

Experiments

Setup

We merge the train and val sets of DOTA-v2.0 [3, 4] into a trainval set, follow the partial-label protocol of Semi-DETR [12] to sample the 10%, 20% and 30% labeled splits from it, and add a fully labeled setting that uses the entire trainval set as labeled data and the test-challenge images as unlabeled data, after we exclude the frames under heavy cloud or fog and the frames whose foreground is almost black. The fully labeled setting starts the unlabeled loss at 20k instead of 10k and is otherwise identical. The detector is DINO [11] with a ResNet-50 [23] backbone. Training runs for 180k iterations with AdamW at a learning rate of 10−410^{-4}, decayed at 120k and 160k. We set the EMA momentum to 0.999, the unsupervised loss weight to 0.5, the pseudo-label score threshold to 0.4, and λreg=5\lambda_{\mathrm{reg}}=5. As in the one-to-many-first design of Semi-DETR [12], the warm-up length equals the full schedule, so every system in Table 1 stays in the one-to-many stage and the comparison is not confounded by the stage switch. The target-selection ablation keeps the legacy 160k pseudo-label DN start and the nested settings use t0=120kt_0=120\mathrm{k} and T=10kT=10\mathrm{k}. All numbers are from the official DOTA evaluation server on test-dev, and we report AP50_{50}, AP75_{75} and mean average precision (mAP). Under the partial-label protocol, results for Oriented Semi-DETR are averaged over three training runs.

Main results

Table 1 compares Oriented Semi-DETR with the DETR-based supervised and semi-supervised baselines at each label ratio. Against the best mAP among them, the gains are 1.78, 2.18, and 2.59 points at 10%, 20%, and 30%; against the best AP75_{75} they are 2.15, 3.39, and 4.04 points. The larger gain at the strict threshold is consistent with better alignment of matched boxes, although aggregate metrics do not separate recall from localization. Table 2 decomposes the full configuration into the three pseudo-label denoising components at every ratio, with all three regression targets swap-aligned throughout. Teacher-score weighting gives the best mAP at 10%. At 20% the encoding margin restores AP50_{50} but reduces AP75_{75}, leaving mAP at the level of the ramp alone. At 30% the full configuration leads on AP50_{50} and mAP. No component is best on every metric at every partial ratio, so we retain the full configuration throughout.

With the entire trainval set used as labeled data, Oriented Semi-DETR reaches 53.16 AP50_{50}, 35.07 AP75_{75} and 33.23 mAP. The same baseline is strongest on all three metrics here, so the gains are 1.45, 2.81 and 2.01 points, widest once more at the strict threshold. The step from the 30% split is about one point of mAP, and two factors limit it. The 180k schedule is the one used for the partial splits and is short for a labeled set three times the size, and the test-challenge images that serve as unlabeled data carry more noise and more hard cases than the trainval set. The three denoising components also rank consistently here, unlike at any partial ratio: the full configuration leads on all three metrics, and the encoding margin recovers the AP75_{75} that teacher-score weighting alone loses.

Ablation

Table 3 isolates the three regression targets. The swap-invariant cost of Equation (4) is active in every row, so the table measures the target selection of Equation (3) alone.

Applying the rule to the supervised DN target alone yields +1.9 points of AP75_{75}. This supports the argument of § 1.2.3 that denoising is the path most exposed to the ambiguity, since the injected noise is small and an encoding mismatch then dominates the correction the decoder must make. Adding the ordinary regression target contributes a further 3.0 points, the largest single step, as expected from the target that carries most of the gradient under one-to-many assignment. The fourth row adds the unsupervised DN target, active over the final 20k iterations. It gains 0.54 AP50_{50} and 0.06 mAP and loses 0.14 AP75_{75}, so the gain appears at the loose threshold alone.

The last three rows add the changes of Section 2.5. Relative to the legacy 160k schedule, the early ramp changes AP75_{75} by +0.17 and mAP by +0.01. Adding the teacher score gives a further +0.03 AP75_{75} and +0.22 mAP. Adding the encoding margin raises AP50_{50} by understates the effect, which is largest while the pseudo-label distribution is being formed.

Method10% labeled20% labeled30% labeledFully labeled
AP50_{50}AP75_{75}mAPAP50_{50}AP75_{75}mAPAP50_{50}AP75_{75}mAPAP50_{50}AP75_{75}mAP
DINO [11]40.7221.7622.4946.3124.9425.8248.1126.7127.1649.1126.9127.51
Semi-DETR [12]43.9425.6225.8248.0128.1028.1050.6329.4729.6051.1131.5030.81
Sparse Semi-DETR [13]44.1625.4925.8348.3828.2828.5150.2429.4029.5351.7132.2631.22
Oriented Semi-DETR45.9227.7727.6149.8931.6730.6952.1133.5132.1953.1635.0733.23

Table 1. The results on DOTA-v2.0 test-dev. Oriented Semi-DETR uses the full configuration of Equations (5)–(7). The last group uses the whole trainval set as labeled data.

Pseudo-label DN setting10% labeled20% labeled30% labeledFully labeled
AP50_{50}AP75_{75}mAPAP50_{50}AP75_{75}mAPAP50_{50}AP75_{75}mAPAP50_{50}AP75_{75}mAP
Early pseudo-label DN ramp45.5627.8827.5249.1932.1530.6951.8433.5731.9652.7934.8633.15
Early ramp + teacher score45.9227.7727.6549.6431.9530.5052.0833.6032.1853.0234.5032.99
Early ramp + teacher score + encoding margin45.9227.7727.6149.8931.6730.6952.1133.5132.1953.1635.0733.23

Table 2. The results on DOTA-v2.0 test-dev with pseudo-label denoising across label ratios. All rows use all three regression targets swap-aligned and successively add the early ramp, teacher-score weighting and encoding-margin weighting.

Regr. targetSup. DNUnsup. DNPseudo-DN scheduleAP50_{50}AP75_{75}mAP
––49.1528.6828.69
✓–50.4730.5530.14
✓✓–51.4433.5431.89
✓✓✓–51.9833.4031.95
✓✓✓early ramp51.8433.5731.96
✓✓✓early ramp + qq52.0833.6032.18
✓✓✓early ramp + qq + mm52.1133.5132.19

Table 3. Contribution of each regression target. The swap-invariant GMM cost is active in all rows, and the last column gives the pseudo-label denoising design. The results are from DOTA-v2.0 test-dev under 30% labeled protocol.

Train stepMatched pairsSigmaAssign. flipFilter flip
20k467116.89%18.88%7.02%
180k38263.89%4.89%2.88%

Table 5. Fixed vs. swap-invariant reliability cost on teacher outputs, 128 unlabeled images. “Sigma” is the fraction of matched pairs whose swapped encoding is closer, and the flip columns count pairs whose assignment or filter decision changes.

Splitting the same pairs by aspect ratio confines the effect to near-square boxes. In the aspect-ratio range (1,1.3](1, 1.3] the swapped encoding is closer for 43.9% of the pairs, 45.5% of the assignments change, and the retention rate rises from 28.6% to 44.9%, while above ratio 2 the first two numbers drop to 7.1% and 8.6%. The overall retention rate therefore shifts by only 1.20 points at 20k, although 7.02% of the individual decisions change; the swap-invariant cost retains a comparable number of pseudo boxes and a different selection of them. Storage tanks, roundabouts and vehicles seen from above make near-square objects common in aerial scenes [3, 4], so a mechanism that would be marginal in general detection is important here.

Prevalence of the ambiguity

Table 5 reports the diagnostic behind the claim in Section 1. We take teacher pseudo boxes and raw student queries from a training checkpoint, run the Hungarian assignment and the GMM filter twice, once with the fixed cost and once with Equation (4), and compare the outcomes.

The fixed matcher selects a swapped query in only 0.09% of the pairs at 20k, against 16.89% for the swap-invariant matcher. Mis-pricing is not the only consequence. The assignment is diverted away from the ambiguous pairs before the filter is applied. Both rates fall as the teacher converges, so a measurement taken late in training

0.03 and mAP by 0.01 but reduces AP75_{75} by 0.09, so this screening does not establish an additional strict-IoU benefit.

Table 4 varies the angle penalty φ\varphi of Equation (2) with everything else fixed at the configuration that aligns all three regression targets. The wrapped form leads the cosine form by 2.15 AP50_{50}, 1.81 AP75_{75} and 1.36 mAP, and raw L1 falls between the two. With λ\lambda held at 1, the cosine row also differs in the effective weight of the angle term, so the table identifies the better configuration rather than isolating the shape. It concerns φ\varphi alone and is therefore reported apart from Table 3.

φ(Δa)\varphi(\Delta a)AP50_{50}AP75_{75}mAP
1−cos⁡(2πΔa)1-\cos(2\pi\Delta a) (cosine)49.8331.5930.59
∣Δa∣|\Delta a| (raw L1)51.5333.0431.76
min⁡(∣Δa∣,1−∣Δa∣)\min(|\Delta a|,1-|\Delta a|)51.9833.4031.95

Table 4. Angle penalty φ\varphi in Equation (2), with all three regression targets swap-aligned. The results are from 30% labeled protocol.

Conclusion

The parameterization of an oriented box is two-to-one: b and σ(b)\sigma(b) denote the same rectangle, so a pseudo label does not determine a unique regression target. Because the training signal is computed from the parameters, the ambiguity acts twice. The regression loss penalizes a student box that coincides with the pseudo box in space, and the Gaussian mixture filter removes that box from the unlabeled loss. Oriented Semi-DETR resolves the ambiguity at both points. It aligns every regression target with the encoding the student already predicts, minimizes over the two encodings inside the reliability cost, and starts pseudo-label denoising earlier under teacher-score and encoding-margin weights. The architecture and the per-iteration computation are unchanged. On DOTA-v2.0 our method improves performance on every label ratio over DETR-based supervised and semi-supervised baselines, and the gain is largest at the strict IoU threshold. The same approach applies wherever a teacher–student target representation admits several equivalent encodings of one oriented object.

Acknowledgements

This work was supported by the Science Fund for Creative Research Groups of the National Natural Science Foundation of China (62521007) and the Joint Key Project of the National Natural Science Foundation of China and China Electronics Technology Group Corporation (U21B2049).

supp.pdf

This document collects the material that did not fit in the main paper: the properties of the two encodings and their proofs, a worked example of the cost inflation, an implementation pitfall in the swap itself, the complete set of ablations we ran, the full diagnostic tables, and the training hyperparameters. Section and equation numbers carry the prefixes A, B, … and S1, S2, … Numbers without a prefix refer to the main paper.

Notation and the angle metric

We write an oriented box as b=(cx,cy,w,h,θ)\mathbf{b} = (c_x, c_y, w, h, \theta) with the angle taken in [−π/2,π/2)[-\pi/2, \pi/2). The center and the size are normalized by the image width and height, and the angle is stored as a=θ/π+1/2∈[0,1)a = \theta/\pi+ 1/2 \in[0, 1). In these coordinates the half-turn of Eq. (1) becomes a↦a+1/2 mod 1a \mapsto a + 1/2 \bmod1, so a swap always moves the normalized angle by exactly 1/21/2.

The distance of Eq. (2) is

d(p,t)=∣Δcx∣+∣Δcy∣+∣Δw∣+∣Δh∣+λφ(Δa),(8)d(\mathbf{p}, \mathbf{t}) = |\Delta c_x| + |\Delta c_y| + |\Delta w| + |\Delta h| + \lambda\varphi(\Delta_a), \tag*{(8)}

and we studied three choices of φ\varphi:

φcos⁡(Δa)=1−cos⁡(2πΔa),(9)\varphi_{\cos}(\Delta_a) = 1 - \cos(2\pi\Delta_a), \tag*{(9)}
φwrap(Δa)=min⁡(∣Δa∣,1−∣Δa∣),(10)\varphi_{\mathrm{wrap}}(\Delta_a) = \min(|\Delta_a|, 1 - |\Delta_a|), \tag*{(10)}
φraw(Δa)=∣Δa∣.(11)\varphi_{\mathrm{raw}}(\Delta_a) = |\Delta_a|. \tag*{(11)}

Table S1 lists their values at the two angles that matter. The first two are π\pi-periodic in θ\theta, so they treat a=0a = 0 and a=1a = 1 as the same angle. The third is not, and it assigns a full unit of penalty across the wrap point. We include it because the angle difference is small most of the time once the target is swap-aligned, which makes the defect less visible than one would expect, but the defect is real.

φ\varphiΔa=1/4\Delta_a = 1/4Δa=1/2\Delta_a = 1/2periodicsmooth
cosine1.002.00yesyes
wrapped L10.250.50yesno
raw L10.250.50nono

Table S1. The three angle penalties. Δa=1/4\Delta_a = 1/4 is a 45° disagreement and Δa=1/2\Delta_a = 1/2 is the swap. λ=1\lambda= 1 in every configuration.

Two remarks on Table S1. First, λ\lambda stays at 1 for all three forms, so the three variants differ in the scale of the angle term as well as in its shape. At a 45° disagreement the cosine penalty is four times the wrapped one. Section F.3 therefore compares three complete configurations, and it leaves the shape of the function uncontrolled. Second, the cosine form has a maximum at the swap, where its derivative vanishes. That stationary point is unstable rather than attracting, so we do not attribute the behavior of the cosine runs to it.

Properties of the two encodings

Write R(b)⊂R2R(\mathbf{b}) \subset\mathbb{R}^{2} for the set of points covered by the rectangle that b\mathbf{b} encodes.

is an involution and the two encodings are distinct

Applying Eq. (1) twice exchanges ww and hh back and adds π\pi to the angle, which the wrap removes, so σ∘σ=id⁡\sigma\circ\sigma= \operatorname{id} and σ\sigma is an involution. The two encodings are always distinct as parameter tuples, including for squares: wrap⁡(θ+π/2)=θ\operatorname{wrap}(\theta+ \pi/2) = \theta has no solution in [−π/2,π/2)[-\pi/2, \pi/2). Hence ∣{b,σ(b)}∣=2|\{\mathbf{b}, \sigma(\mathbf{b})\}| = 2 for every b\mathbf{b}.

preserves the rectangle

Let u(θ)=(cos⁡θ,sin⁡θ)\mathbf{u}(\theta) = (\cos\theta, \sin\theta) and v(θ)=(−sin⁡θ,cos⁡θ)\mathbf{v}(\theta) = (-\sin\theta, \cos\theta) be the two edge directions, so that

R(b)={(cx,cy)+su(θ)+tv(θ):∣s∣≤w2, ∣t∣≤h2}.(12)R(\mathbf{b}) = \left\{(c_x,c_y) + s\mathbf{u}(\theta) + t\mathbf{v}(\theta) : |s| \le\frac{w}{2},\ |t| \le\frac{h}{2}\right\}. \tag*{(12)}

A quarter turn gives u(θ+π/2)=v(θ)\mathbf{u}(\theta+ \pi/2) = \mathbf{v}(\theta) and v(θ+π/2)=−u(θ)\mathbf{v}(\theta+ \pi/2) = -\mathbf{u}(\theta). Substituting these into Eq. (12) for σ(b)\sigma(\mathbf{b}), whose half-extents are h/2h/2 along u(θ+π/2)\mathbf{u}(\theta+\pi/2) and w/2w/2 along v(θ+π/2)\mathbf{v}(\theta+\pi/2), reproduces the same point set, since the range of ss and tt is symmetric about zero. The wrap adds a multiple of π\pi when the result leaves [−π/2,π/2)[-\pi/2,\pi/2), and a half turn sends u↦−u\mathbf{u} \mapsto-\mathbf{u} and v↦−v\mathbf{v} \mapsto-\mathbf{v}, which again leaves Eq. (12) unchanged. Therefore R(σ(b))=R(b)R(\sigma(\mathbf{b})) = R(\mathbf{b}).

The map from parameters to rectangles is consequently two-to-one, and its fibers are exactly the pairs {b,σ(b)}\{\mathbf{b},\sigma(\mathbf{b})\}. An annotation protocol selects one member per fiber and can do so consistently. A teacher network has no such constraint, which is the source of the problem the main paper addresses.

The selection never increases the regression distance

Let d0=d(p,t)d_0 = d(\mathbf{p}, \mathbf{t}) and d1=d(p,σ(t))d_1 = d(\mathbf{p}, \sigma(\mathbf{t})). By definition t∗\mathbf{t}^* of Eq. (3) attains min⁡(d0,d1)\min(d_0,d_1), so

d(p,t∗)=min⁡(d0,d1)≤d0,(13)d(\mathbf{p}, \mathbf{t}^*) = \min(d_0,d_1) \le d_0, \tag*{(13)}

with equality whenever the teacher’s encoding is already the closer member. The selection is therefore never worse than the ordinary target on the quantity the loss measures, and it is inactive on the majority of pairs. Table S7 gives the fraction on which it is active.

StepPairsSigma fixedSigma swapAssign. flipFilter flipKeep fixedKeep swapSwap onlyFixed only
20k46710.09%16.89%18.88%7.02%35.54%36.74%4.11%2.91%
180k38260.05%3.89%4.89%2.88%36.23%36.02%1.33%1.54%

Table S7. Fixed versus swap-invariant reliability cost on teacher outputs from 128 unlabeled images. “Sigma” is the fraction of matched pairs whose swapped encoding is closer, under each matcher. “Assign. flip” counts pairs whose matched query changes, “Filter flip” counts pairs whose keep or drop decision changes. “Keep” is the fraction of pseudo boxes the GMM retains. The last two columns split the filter flips into boxes kept only under the swap-invariant cost and boxes kept only under the fixed cost.

The training signal no longer depends on the encoding

Since {σ(t),t}={t,σ(t)}\{\sigma(\mathbf{t}), \mathbf{t}\} = \{\mathbf{t}, \sigma(\mathbf{t})\}, replacing the teacher output t\mathbf{t} by σ(t)\sigma(\mathbf{t}) leaves the feasible set of Eq. (3) unchanged and therefore leaves t∗\mathbf{t}^* unchanged. The loss the student receives is invariant under the teacher’s choice of encoding. This is the property we are after, and no reweighting of the two candidates achieves it, because a weighted target still moves when the weights are applied to a different member.

Why a hard arg min

Averaging the two encodings is not viable, because the pair is not convex in parameter space. Take b=(0,0,4,1,0)\mathbf{b} = (0,0,4,1,0), so that σ(b)=(0,0,1,4,π/2)\sigma(\mathbf{b}) = (0,0,1,4,\pi/2). The midpoint of the two is (0,0,2.5,2.5,π/4)(0,0,2.5,2.5,\pi/4), a square of area 6.25 turned by 45∘45^\circ. It shares neither the shape nor the area of the original rectangle, whose area is 4. Any soft combination of the two encodings gives a box that is not the annotated object.

Cost of the rule

t∗\mathbf{t}^* is computed from sg⁡(p)\operatorname{sg}(\mathbf{p}), so ∂t∗/∂p=0\partial\mathbf{t}^*/\partial\mathbf{p} = 0 and the backward pass is that of the ordinary loss with a constant target. The rule introduces no parameter, no additional forward pass and no change to the architecture. Its runtime cost is one extra evaluation of Eq. (8) and one comparison per target. Inference is unchanged.

A worked example of the cost inflation

Section 2.4 states that the fixed regression cost reaches about 11.5 for a perfectly localized box. The arithmetic is as follows. Take a pseudo box whose normalized width and height are w=0.25w = 0.25 and h=0.10h = 0.10, which is a 256 ×\times 102 pixel object in a 1024 ×\times 1024 crop, and a student query that predicts

exactly σ(b)\sigma(\mathbf{b}). The two boxes are the same rectangle, so the student has localized the object without error. The individual terms of the fixed cost are listed in Table S2.

TermFixed cost
∣Δcx∣+∣Δcy∣|\Delta c_x| + |\Delta c_y|0
∣Δw∣+∣Δh∣|\Delta w| + |\Delta h|0.30
λθcos⁡(Δθ)\lambda_{\theta}\cos(\Delta_{\theta}), Δα=1/2\Delta_{\alpha}=1/22.00
dd2.30
Creg=5dC_{\mathrm{reg}} = 5d11.5

Table S2. Term-by-term breakdown of the fixed Hungarian regression cost for a student query that predicts the swapped encoding of a perfectly localized pseudo box. The cost weight is 5 and λ=1\lambda= 1, matching our configurations.

The angle term alone contributes 5(1−cos⁡π)=105(1-\cos\pi)=10, and the exchanged size adds 5×0.30=1.55 \times0.30 = 1.5. The rotated IoU term of the same matching cost treats the two encodings as one rectangle and reports a perfect overlap, so the fixed cost is internally inconsistent: one of its terms reports an exact match and another reports the worst possible one. Eq. (4) returns 0 and removes the contradiction.

The consequence for the filter is not confined to the ambiguous pairs. Semi-DETR [1] fits a two-component GMM to the matched costs of each iteration and keeps the pseudo boxes assigned to the low-cost component. A subset of costs inflated by roughly 11.5 forms a heavy right tail, which pulls the fitted components apart and moves the decision boundary for every pseudo box in the batch. This is why Table S7 shows a large number of filter flips while the total number of retained boxes changes only slightly.

The swap under anisotropic normalization

The exchange of ww and hh must be carried out on absolute pixel values. Doing it on the normalized values is wrong whenever the image is not square, and the failure mode is subtle enough to deserve a note.

Let the image be W×HW \times H and write wˉ=w/W\bar{w} = w/W and hˉ=h/H\bar{h} = h/H. Swapping the normalized values gives back the pixel sizes

w′=hˉW=hWH,h′=wˉH=wHW.(14)w' = \bar{h}W = h\frac{W}{H}, \qquad h' = \bar{w}H = w\frac{H}{W}. \tag*{(14)}

The intended result is w′=hw' = h and h′=wh' = w, so Eq. (14) is off by a factor W/HW/H on one side and H/WH/W on the other. The aspect ratio of the result is the intended one multiplied by (W/H)2(W/H)^2. Table S3 works through a case with W/H=2W/H = 2, where a 400 ×\times 100 box turns into a square instead of a 100 ×\times 400 box.

wˉ\bar{w}hˉ\bar{h}pixels
input0.39060.1953400×100400 \times100
correct swap0.09770.7813100×400100 \times400
normalized swap0.19530.3906200×200200 \times200

Table S3. Swapping the normalized size instead of the pixel size, for a 400×100400 \times100 pixel box in a 1024×5121024 \times512 image. Both results have the same area as the input.

Note that w′h′=whw'h' = wh holds for Eq. (14) at every aspect ratio, so the area of the box is preserved and a check on area or on scale will not detect the mistake. Only the aspect ra-

Algorithm S1 Swap-aligned targets in one iteration

Require: student predictions {pi}\{p_i\}, targets {tj}\{t_j\}, distance dd

  1. run the teacher, threshold its scores, obtain pseudo boxes

  2. assign queries to targets with the Hungarian matcher using Eq. (4) as the regression cost

  3. fit the two-component GMM to the matched costs, keep the reliable set

  4. for each matched pair (pi,tj)(p_i,t_j) and each DN pair do

  5. tjsw←σ(tj)t_j^{\mathrm{sw}} \leftarrow\sigma(t_j) {exchange on pixels, then renormalize}

  6. d0←d(sg⁡(pi),tj)d_0 \leftarrow d(\operatorname{sg}(p_i),t_j)

  7. d1←d(sg⁡(pi),tjsw)d_1 \leftarrow d(\operatorname{sg}(p_i),t_j^{\mathrm{sw}})

  8. tj∗←tjt_j^* \leftarrow t_j if d0≤d1d_0 \le d_1, else tjswt_j^{\mathrm{sw}}

  9. end for

  10. accumulate the losses against {tj∗}\{t_j^*\} and update the student

  11. update the teacher by EMA

ratio changes. The rule is active mostly on near-square targets, where the correct swap is close to the identity, so the error replaces an almost correct target by a clearly wrong one while every aggregate size statistic remains unchanged. Aerial crops are often square, which hides the mistake further. We swap on pixel values and renormalize.

Algorithm

Algorithm S1 states the rule as it is applied inside one training iteration. Lines 4 to 8 run on the three regression targets of Section 2.3, once for the regression targets of the queries selected by the one-to-many assigner, once for the supervised DN targets and once for the pseudo-label DN targets. Line 11 is the reliability cost of Eq. (4). The teacher is unchanged, and so is inference.

The tie d0=d1d_0 = d_1 is resolved in favor of the original encoding, which keeps the selection inactive when it would not reduce the distance. The comparison returns an index rather than a differentiable quantity, so the selection creates no gradient path.

Complete set of ablations

Table S4 lists every configuration we trained to completion on the 30% split. Each row is averaged over three training runs, except the fixed-reliability-cost control, which was trained once. All runs use the same detector, the same 180k schedule and the same pseudo-label threshold, and all of them keep the one-to-many warm-up equal to the total schedule, so none of the comparisons below is confounded by the Semi-DETR [1] stage switch.

Three groups of runs appear in the table. The rows marked cosine are an earlier sequence, which covers the full 2×22 \times2 over the two denoising paths and allows us to attribute the gain one path at a time. The rows marked wrapped are the sequence reported in the main paper. The two sequences differ in the angle penalty φ\varphi and cannot be pooled. The final four configurations retain wrapped L1 and align all three regression targets, then vary only pseudo-label denoising. The configuration with an early ramp, teacher-score weighting, and encoding-margin weighting is the form of Oriented Semi-DETR reported in the main paper.

The reliability cost on its own

The fixed-reliability-cost and swap-invariant-cost controls differ only in the regression term of the Hungarian cost, with target selection disabled everywhere. The difference is −0.12-0.12 AP50_{50}, +0.35+0.35 AP75_{75}, and −0.03-0.03 mAP. The swap-invariant cost gains at the strict threshold but loses marginally at the loose threshold, so this control does not establish a final-accuracy gain by itself. The comparison is unbalanced because the fixed-cost control was run once and the swap-invariant-cost control three times. Both operate near 42 AP50_{50}, far below the 50.63 reached by Semi-DETR [1], because the metric control replaces plain L1 by the π\pi-periodic cosine form everywhere while leaving every target unchanged. That substitution alone costs about eight points, so the pair probes a weak operating point.

This is the reason the main paper presents Eq. (4) as mechanism evidence, supported by the diagnostic of Section H, and keeps it enabled in every row of Table 2 rather than claiming an AP gain for it.

Which denoising path matters

The cosine configurations with a swap-invariant cost and regression target form a 2×22 \times2 study over the two DN paths. Adding supervised DN gives +6.69+6.69 AP50_{50} and +6.79+6.79 AP75_{75} over the regression-target-only configuration. Pseudo-label DN alone changes these metrics by −0.40-0.40 and −0.64-0.64 points. Adding it on top of supervised DN changes them by −0.51-0.51 and −0.74-0.74 points, respectively.

The wrapped-L1 configurations lead to the same qualitative conclusion. Adding supervised DN to the swap-invariant cost raises AP75_{75} by 1.87 points, and adding the regression target raises it by another 2.99 points. Adding pseudo-label DN target selection changes AP75_{75} by −0.14-0.14 points. Because the two angle penalties were trained in different batches, these values are not pooled. The pseudo-label DN path remains enabled because the wrapped-L1 configuration with all three regression targets aligned improves AP50_{50} by 0.54 points and mAP by 0.06 points over the configuration without that path. It is active only during the last 20k iterations of the 180k schedule. The early-ramp study below lengthens this window.

Choice of the angle penalty

The cosine, wrapped-L1 and raw-L1 configurations that align all three regression targets differ only in φ\varphi. Relative to the cosine form, wrapped L1 gains 2.15 AP50_{50}, 1.81 AP75_{75}, and 1.36 mAP; raw L1 gains 1.70, 1.45, and 1.17 points. The ordering is the same for all three metrics, so we use wrapped L1.

The caveat of Section A applies in full. Since λ\lambda stays at 1, the three rows differ in the effective weight of the angle term as well as in the shape of the penalty, and the cosine form assigns four times as much as the other two at a 45° disagreement. The comparison identifies the better configuration; it does not isolate the function shape. A clean shape study would rescale λ\lambda so that the three penalties agree at a chosen reference angle, and we have not run it.

The cosine configuration with angle-aware one-to-many alignment adds 0.61 AP50_{50}, 0.49 AP75_{75}, and 0.38 mAP over the corresponding cosine configuration. We did not rerun this addition under wrapped L1, and it is not part of Oriented Semi-DETR.

A representation-level alternative

The double-angle-head configuration replaces the scalar angle output by the continuous vector (cos⁡2θ,sin⁡2θ)(\cos2\theta, \sin2\theta), decoded back to the same five-parameter interface. Apart from that change, it matches the wrapped-L1 configuration with supervised DN. It gains +0.02 AP50_{50}, +0.29 AP75_{75}, and +0.25 mAP.

The outcome is what the analysis of Section B predicts. The double-angle vector removes the discontinuity of the scalar angle at the wrap point, since θ\theta and θ+π\theta+\pi map to the same vector. It does not remove the swap ambiguity, because σ\sigma negates the vector and exchanges ww and hh at the same time. The ambiguity is a joint property of (w,h,θ)(w,h,\theta), so no re-encoding of the angle alone can be invariant under it. This separates our rule from the angle encoding literature, CSL [2] and PSC [3] among others, and it is also why the Gaussian losses of § 1.3 are the relevant point of comparison: a Gaussian is a function of the rectangle itself. The head was not recalibrated after this change, so this is not a conclusive comparison of output representations.

The pseudo-label denoising schedule

The final four configurations retain the aligned wrapped-L1 design and change only how pseudo-label denoising is opened and weighted. The first moves the start from 160k to 120k and adds a 10k linear ramp, with no per-query weight. The second multiplies the regression weight of each denoising query by the teacher score qq of its pseudo box. The third additionally multiplies the angle term by

m=∣d(p,b)−d(p,σ(b))∣d(p,b)+d(p,σ(b))+ε(15)m = \frac{\lvert d(\mathbf{p}, \mathbf{b}) - d(\mathbf{p}, \sigma(\mathbf{b})) \rvert}{d(\mathbf{p}, \mathbf{b}) + d(\mathbf{p}, \sigma(\mathbf{b})) + \varepsilon} \tag*{(15)}

with ε=10−6\varepsilon= 10^{-6}, which is large when the two encodings are clearly distinguishable and small when the prediction lies between them. The fourth configuration redistributes angle weight by the target aspect ratio r=max⁡(w,h)/min⁡(w,h)r = \max(w,h)/\min(w,h), using the mass-conserving multiplier u=1+(r2−1)/(r2+1)u = 1 + (r^2 - 1)/(r^2 + 1) normalized to keep the total angle weight fixed.

The encoding-margin and aspect-ratio controls do not establish a uniform gain over teacher-score weighting. Both reduce the angle emphasis for near-square targets. For aspect-ratio weighting, this follows directly because uu increases with rr. The diagnostic in Section H shows the opposite distribution for the ambiguity itself: at 180k it occurs for 22.1% of near-square targets and 1.2% of elongated targets. Aspect-ratio weighting therefore transfers angle supervision away from the group where the ambiguity is most common. Because the re-distribution conserves total mass, this transfer is not offset elsewhere.

Averaged over pseudo-label DN positives, the encoding margin is 0.855, so it reduces angle weight by about six percent relative to spatial weight. The aspect-ratio multiplier is bounded in [1,2)[1, 2) before normalization and produces about a 1.5× tilt between near-square and elongated targets on the observed distribution. These controls are retained as component analyses rather than presented as separate methods.

Results across label ratios

The main paper reports the same DOTA-v2.0 [4, 5] partial-label protocol of Semi-DETR [1] at 10%, 20%, and 30% labeled images. Table S5 gives the mean test-dev metrics for DINO [6], Semi-DETR [1], Sparse Semi-DETR [7], and Oriented Semi-DETR. The three component settings below the main methods are controlled parts of Oriented Semi-DETR, and we do not name them as separate methods. They successively add an early pseudo-label DN ramp, teacher-score weighting, and encoding-margin weighting.

Method or component setting10% labeled AP₅₀10% labeled AP₇₅10% labeled mAP20% labeled AP₅₀20% labeled AP₇₅20% labeled mAP30% labeled AP₅₀30% labeled AP₇₅30% labeled mAPFully labeled AP₅₀Fully labeled AP₇₅Fully labeled mAP
DINO40.7221.7622.4946.3124.9425.8248.1126.7127.1649.1126.9127.51
Semi-DETR43.9425.6225.8248.0128.1028.1050.6329.4729.6051.1131.5030.81
Sparse Semi-DETR44.1625.4925.8348.3828.2828.5150.2429.4029.5351.7132.2631.22
Oriented Semi-DETR45.9227.7727.6149.8931.6730.6952.1133.5132.1953.1635.0733.23
Early pseudo-label DN ramp45.5627.8827.5249.1932.1530.6951.8433.5731.9652.7934.8633.15
Early ramp + teacher score45.9227.7727.6549.6431.9530.5052.0833.6032.1853.0234.5032.99
Early ramp + teacher score + encoding margin45.9227.7727.6149.8931.6730.6952.1133.5132.1953.1635.0733.23

Table S5. Mean DOTA-v2.0 test-dev performance across label ratios and in the fully labeled setting. Oriented Semi-DETR uses the early ramp, teacher score, and encoding margin.

At 10%, teacher-score weighting gives the highest mAP (27.65), the early ramp gives the highest AP75_{75} (27.88), and teacher-score and encoding-margin weighting tie in AP50_{50} (45.92). At 20%, encoding-margin weighting has the highest AP50_{50} (49.89), the early ramp has the highest AP75_{75} (32.15), and the early-ramp and encoding-margin configurations tie in mAP (30.69). At 30%, the corresponding values are already reported in the main ablation. In the fully labeled setting the ordering is cleaner: the complete configuration has the highest AP50_{50} (53.16), AP75_{75} (35.07) and mAP (33.23), and the encoding margin recovers the AP75_{75} that teacher-score weighting alone loses (34.50 against 34.86 for the ramp alone). Across the partial splits, however, no component setting dominates every metric at every label ratio. We re-port the nested component study explicitly and use Oriented Semi-DETR only for the complete method.

Moving from the 30% split to the fully labeled setting yields about one point of mAP, less than the step from 20% to 30%. Two properties of the protocol limit it. The 180k schedule is inherited unchanged from the partial splits and is short for a labeled set three times the size, so the extra annotation is seen fewer times per image. The unlabeled pool also changes character: the test-challenge images carry more noise and more hard cases than the training set, even after we drop the frames under heavy cloud or fog and the frames whose foreground is almost black. A longer schedule and a cleaner unlabeled pool are the two direct ways to test whether the gap is a property of the method or of the budget, and we have run neither.

The diagnostic in full

Protocol

We load a training checkpoint, run the teacher on 128 unlabeled images and keep the pseudo boxes above the usual score threshold. The raw student queries of the same forward pass are then matched to those pseudo boxes twice. The first pass uses the fixed regression cost, the second uses Eq. (4). The classification and rotated IoU terms are shared between the two passes, and each pass runs its own Hungarian assignment and fits its own two-component GMM to the matched costs of that batch, exactly as training does. A record is one matched pseudo box, so every percentage below is taken over matched pairs. We report two checkpoints, 20k and 180k, to show the effect at the start and at the end of training.

Overall behavior

Table S7 gives the complete counts. “Sigma” is the fraction of matched pairs for which the swapped encoding of the pseudo box is closer to the matched query than the teacher’s own encoding. It is reported for both matchers, and the gap between the two columns is the first result: the fixed matcher selects a swapped query on its own for 0.09% of the pairs at 20k, against 16.89% for the swap-invariant matcher. Beyond mis-pricing the ambiguous pairs, the fixed cost diverts the assignment away from them, so the filter never evaluates them. This is also why the diagnostic must rerun the assignment rather than only rescoring a fixed matching.

Class10% labeled Early ramp10% labeled + Teacher score10% labeled + Encoding margin20% labeled Early ramp20% labeled + Teacher score20% labeled + Encoding marginFully labeled Early rampFully labeled + Teacher scoreFully labeled + Encoding margin
plane77.0177.1876.9277.1978.6578.8879.4179.1678.98
baseball-diamond54.5654.6153.7853.5552.2653.8549.6250.5352.92
bridge45.3745.0444.3648.6748.7748.6248.7748.6349.01
ground-track-field61.9262.4262.7465.8264.6665.4367.1767.9067.28
small-vehicle36.2435.1936.3539.3039.6738.0041.4140.8340.88
large-vehicle41.2241.7340.9942.4443.8043.7145.2344.9745.72
ship44.1343.5443.5044.5447.9346.7850.9450.9751.01
tennis-court76.4076.3477.2074.4974.1873.6682.6582.4582.26
basketball-court54.7553.3354.8558.1357.2756.1163.1862.4862.60
storage-tank59.2459.1459.1259.7259.6259.5063.5362.3862.28
soccer-ball-field32.4834.5032.8439.6341.0040.3743.4345.9745.61
roundabout50.4050.8549.9353.0852.5252.6257.5257.1057.98
harbor39.6439.4539.7338.3238.6038.3643.9643.9044.16
swimming-pool50.3151.1050.5051.8052.6452.3953.8453.9953.98
helicopter37.2338.8242.1941.6247.0751.1358.3260.0258.85
container-crane4.377.597.3212.6211.6614.027.297.758.80
airport47.8750.2150.2878.0676.7176.7185.6884.4985.67
helipad6.875.604.016.426.497.828.3010.898.86

Table S6. Class-wise AP₅₀ means for the three Oriented Semi-DETR component settings at 10% and 20% labeled images and in the fully labeled setting. The official output provides only class-wise AP₅₀.

The second result is the contrast between the flip columns and the keep columns. At 20k the net keep rate rises by 1.20 points while 7.02% of the individual decisions change, split almost evenly into 4.11% gained and 2.91% lost. At 180k the net keep rate falls by 0.21 points and 2.88% of the decisions still change. The swap-invariant cost retains a comparable number of pseudo boxes and a different selection of them. Any argument for it therefore must concern which boxes it keeps.

Restricting attention to the pairs that the swap-invariant matcher assigns to a swapped query makes the effect more explicit. There are 789 such pairs at 20k, of which the fixed pipeline keeps 1.39% and the swap-invariant pipeline keeps 16.73%. At 180k there are 149, kept at 1.34% and 12.75%. These are the pairs on which a perfectly localized student prediction is priced as a failure, and the fixed pipeline discards almost all of them.

Both rates fall by roughly a factor of four between 20k and 180k. The teacher becomes more consistent as it converges, so the ambiguity is a property of the early pseudo-label distribution more than of the converged one. A measurement taken on a finished model therefore understates it, since it is largest while the pseudo-label distribution is being formed and has mostly disappeared by the time training ends.

Aspect ratio

Table S8 splits the same pairs by the aspect ratio of the pseudo box at both checkpoints. The effect is confined to near-square boxes at both. In the aspect-ratio range (1, 1.3] at 20k, the swapped encoding is closer for 43.92% of the pairs, 45.49% of the assignments change, and the keep rate rises from 28.63% to 44.90%, a gain of 16.27 points. In the range (2, ∞) the first two numbers fall to 7.05% and 8.62% and the keep rate moves by 1.49 points in the other direction. At 180k the near-square range still shows 22.09% sigma and 11.64% filter flips, but the keep rates have converged, so what remains is a substitution of one set of pseudo boxes for another.

StepAspect rationnSigmaAssign. flipFilter flipKeep fixed / swap
20k(1,1.3](1, 1.3]51043.92%45.49%18.63%28.63% / 44.90%
(1.3,2](1.3, 2]128328.22%31.33%8.26%27.36% / 28.60%
(2,∞)(2, \infty)28787.05%8.62%4.41%40.41% / 38.92%
180k(1,1.3](1, 1.3]33522.09%25.97%11.64%41.49% / 41.19%
(1.3,2](1.3, 2]11094.24%5.50%3.16%30.21% / 29.58%
(2,∞)(2, \infty)23821.18%1.64%1.51%38.29% / 38.29%

Table S8. The diagnostic split by aspect ratio at both checkpoints. “Keep” is given under the fixed and the swap-invariant cost.

This is the expected shape. For an elongated box the two encodings differ strongly in ww and hh, so the size terms of Eq. (8) separate them and the teacher’s encoding is almost always the closer one. For a near-square box the size terms nearly cancel and the decision rests on the angle alone. Near-square objects are common in aerial imagery: storage tanks, roundabouts, swimming pools and vehicles seen from above, all represented among the DOTA categories [4, 5]. A mechanism that would be marginal in general object detection is important here.

Class composition

Table S9 reports the five classes that contribute most pairs. The clearest case is the swimming pool, which is close to square in plan view: at 20k the swapped encoding is closer for 57.07% of its pairs and the keep rate rises from 21.20% to 51.63%. Elongated classes behave as Table S8 predicts. Large vehicles show 7.06% sigma at 20k and 0.79% at 180k, and their keep rate changes only slightly.

StepClassnnSigmaAssign. flipFilter flipKeep fixed / swap
20kship3234.38%37.50%9.38%37.50% / 46.88%
small vehicle20016.00%16.50%5.00%55.00% / 56.00%
large vehicle6807.06%8.97%1.62%44.41% / 43.38%
swimming pool18457.07%60.33%33.70%21.20% / 51.63%
storage tank217313.99%16.43%6.26%36.45% / 34.70%
180kship313.23%3.23%3.23%48.39% / 51.61%
small vehicle942.13%2.13%4.26%53.19% / 53.19%
large vehicle5080.79%0.98%0.59%43.70% / 43.50%
swimming pool22526.22%31.11%15.56%38.22% / 36.89%
storage tank20062.09%3.24%2.39%32.75% / 32.45%

Table S9. The diagnostic for the five classes contributing the most matched pairs.

What the diagnostic does not show

Four limits are worth stating explicitly.

The sample is 128 unlabeled images. Pairs from the same image are strongly correlated, so the counts of 4671 and 3826 cannot be treated as independent observations and we do not attach confidence intervals to any of these rates.

The class composition is uneven. Storage tanks alone are 46.5% of the pairs at 20k and 52.4% at 180k, and 59.7% of the near-square pairs at 180k are swimming pools. Aspect ratio and class are therefore confounded, and Tables S8 and S9 should be read as two views of the same population rather than as independent evidence.

The evaluation server returns per-class results at AP50_{50} only. We cannot connect the classes of Table S9 to per-class AP75_{75}, so the diagnostic cannot be checked class by class against the accuracy gain.

Finally, the diagnostic measures routing and filtering at two fixed checkpoints. It identifies the error the fixed cost makes and how often it occurs; it does not by itself establish a gain in final accuracy. The accuracy claims rest on the training runs of Table S4.

ConfigurationReg.Sup. DNUnsup. DNφ\varphiAP50_{50}AP75_{75}mAP
Fixed reliability costcosine42.4523.6024.14
Swap-invariant costcosine42.3323.9524.11
Swap-invariant cost + regression target✓cosine43.6525.5425.50
Swap-invariant cost + regression target + supervised DN✓✓cosine50.3432.3331.20
Swap-invariant cost + regression target + pseudo-label DN✓✓cosine43.2524.9025.07
All three regression targets✓✓✓cosine49.8331.5930.59
All three + angle-aware one-to-many alignment✓✓✓cosine50.4432.0830.97
Swap-invariant cost, wrapped L1wrapped49.1528.6828.69
Wrapped cost + supervised DN✓wrapped50.4730.5530.14
Wrapped cost + supervised DN + regression target✓✓wrapped51.4433.5431.89
All three regression targets, wrapped L1✓✓✓wrapped51.9833.4031.95
All three regression targets, raw L1✓✓✓raw51.5333.0431.76
Wrapped cost + supervised DN + double-angle head✓wrapped50.4930.8430.39
All three + early pseudo-label DN ramp✓✓✓wrapped51.8433.5731.96
Early ramp + teacher score✓✓✓wrapped52.0833.6032.18
Early ramp + teacher score + encoding margin✓✓✓wrapped52.1133.5132.19
Early ramp + teacher score + encoding margin + aspect ratio✓✓✓wrapped51.5133.0231.73

Table S4. Complete ablations on DOTA-v2.0 with 30% labeled images. “Reg.”, “Sup. DN”, and “Unsup. DN” mark the regression targets to which Eq. (3) is applied. All configurations except the fixed-reliability-cost control use the swap-invariant reliability cost of Eq. (4). The early-ramp, teacher-score, and encoding-margin configuration is Oriented Semi-DETR in the main paper.

Hyperparameters

Table S10 lists the settings shared by the runs in Table S4. They use the Semi-DETR-style training pipeline [1]; the shared entries were not tuned for this work. The only configuration-specific exception is the pseudo-label DN schedule, which is reported separately below because the legacy controls and the complete Oriented Semi-DETR configuration use different start times.

The warm-up length deserves a comment because it decides which Semi-DETR [1] stage the runs are in. Semi-DETR uses a one-to-many assignment first and switches to one-to-one after the warm-up. In our protocol, we set the warm-up to 180k, equal to the total schedule, so every run stays in the first stage and the one-to-one branch, together with the cross-view query consistency associated with that stage, is not activated. This is our controlled experimental choice, rather than a claim about a universal Semi-DETR default; it keeps the comparison in Table S4 on the same stage. Consequently, the results here speak to the one-to-many stage, while the effect of the proposed rule in the one-to-one stage remains untested.

BackboneResNet-50 [8], ImageNet initialized
DetectorDINO [6], 900 queries, 100 DN queries
OptimizerAdamW, lr 10−410^{-4}, weight decay 10−410^{-4}
Backbone lr0.1×0.1 \times the base rate
Schedule180k iterations, lr ÷10\div10 at 120k and 160k
Batch1 labeled and 1 unlabeled image per GPU
Train scalelong side 1333, short side sampled in [1024, 1200]
Test scale1024 × 1024
Augmentationrandom flip and rotation, weak for the teacher
TeacherEMA, momentum 0.999
Pseudo labelsscore threshold 0.4
Unsupervised weight0.5
O2M warm-up180k, equal to the schedule
Pseudo-label DN (legacy)160k start
Pseudo-label DN (complete method)120k start + 10k ramp
Matching costfocal 2, regression 5, rotated IoU 2
Lossesbox 5, rotated IoU 2, λ=1\lambda= 1
Seed42

Table S10. Training settings. All entries are shared unless a configuration-specific pseudo-label DN schedule is stated explicitly.

Protocol and open points

All experiments use DOTA-v2.0 [4, 5] and are evaluated by the official server on test-dev. The 10%, 20% and 30% splits follow the partial annotation protocol of Semi-DETR [1], fold 1, sampled from a trainval set that merges the official train and val sets. The fully labeled setting keeps the entire trainval set as labeled data and uses the test-challenge images as unlabeled data, after we drop the frames under heavy cloud or fog and the frames whose foreground is almost black. It starts the unlabeled loss at 20k instead of 10k, and every other hyperparameter, including the 180k schedule, the one-to-many warm-up length and the pseudo-label threshold, is the one used for the partial splits.

Every run uses seed 42, so the runs behind each mean differ only through nondeterminism that the seed does not control, mainly nondeterministic CUDA kernels and data loading order. The means we report therefore average out the nondeterminism that remains at a fixed seed. They do not support a significance test. A multi-seed study is the appropriate way to strengthen the statistical claims, and we have not run one.

Two points remain open. The isolation of the reliability cost rests on an unbalanced comparison between the fixed-reliability-cost control and the swap-invariant-cost control, and would be worth rerunning on equal footing. The angle penalty comparison of Section F.3 confounds the shape of φ\varphi with its scale, and a rescaled control would separate them.

References

References

  1. [1]Jiacheng Zhang, Xiangru Lin, Wei Zhang, Kuo Wang, Xiao Tan, Junyu Han, Errui Ding, Jingdong Wang, and Guanbin Li, “Semi-DETR: Semi-supervised object detection with detection transformers,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 23809–23818.arxiv.org/abs/2307.08095
  2. [2]Xue Yang and Junchi Yan, “Arbitrary-oriented object detection with circular smooth label,” in European Conference on Computer Vision, 2020, pp. 677–694.DOI
  3. [3]Yi Yu and Feipeng Da, “Phase-shifting coder: Predicting accurate orientation in oriented object detection,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 13354–13363.arxiv.org/abs/2211.06368
  4. [4]Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang, “DOTA: A large-scale dataset for object detection in aerial images,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018, pp. 3974–3983.arxiv.org/abs/1711.10398
  5. [5]Jian Ding, Nan Xue, Gui-Song Xia, Xiang Bai, Wen Yang, Michael Ying Yang, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang, “Object detection in aerial images: A large-scale benchmark and challenges,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 11, pp. 7778–7796, 2022.DOI
  6. [6]Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni, and Heung-Yeung Shum, “DINO: DETR with improved denoising anchor boxes for end-to-end object detection,” in International Conference on Learning Representations, 2023.
  7. [7]Tahira Shehzadi, Khurram Azeem Hashmi, Didier Stricker, and Muhammad Zeshan Afzal, “Sparse semi-DETR: Sparse learnable queries for semi-supervised object detection,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 5840–5850.arxiv.org/abs/2404.01819
  8. [8]Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, “Deep residual learning for image recognition,” in IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 770–778.arxiv.org/abs/1512.03385

Paper details

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

views · 0

downloads · 0

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

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents