acceptodds
Withdrawn from ICLR 2027

Imbalance-Adaptive Cross Consistency Learning for Semi-Supervised Oriented Object Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Oriented object detection (OOD) has achieved significant progress under fully supervised learning. However, its performance heavily relies on large-scale annotated datasets, which are costly and time-consuming to obtain. To alleviate this limitation, semi-supervised oriented object detection (SSOOD) has attracted increasing attention due to its ability to leverage unlabeled data. Nevertheless, existing SSOOD methods often suffer from severe class imbalance in pseudo-label generation, particularly under long-tailed data distributions, which hinders performance. To address this issue, we propose Imbalance-Adaptive Cross Consistency Learning (IACCL), a novel SSOOD framework that simultaneously mitigates class imbalance and enhances model robustness. Specifically, we introduce a Balanced Pseudo-Label (BPL) strategy that adaptively regulates the number of pseudo-labels across categories based on their comprehensive scores, thereby mitigating the dominance of head classes and promoting a more balanced learning process. In addition, we design a Cross Consistency Constraint (CCC) to enforce consistency between the teacher and student models under shared feature representations, encouraging the extraction of more robust and discriminative features, thereby alleviating overfitting under limited supervision. Extensive experiments on benchmark datasets demonstrate that the proposed method achieves consistent improvements over the state-of-the-art methods, particularly for tail categories, validating its effectiveness in handling class imbalance in SSOOD.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Remote sensing, Oriented object detection, Semi-supervised learning, Long tail, Cross Consistency Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_RzvCOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes IACCL for semi-supervised oriented object detection, introduces Balanced Pseudo-Label (BPL) to adaptively suppress pseudo-labels from dominant classes using classification confidence and localization quality, and Cross Consistency Constraint (CCC) to supervise teacher and student heads using shared pseudo-labels. Experimental studies on DOTA v1.0 and v1.5 evaluate different labeling ratios, component ablations, and sensitivity to key parameters.

Critical strengths

BPL addresses class imbalance using both classification confidence and localization quality, with category-wise analyses supporting its effectiveness. Also, experiments cover two standard benchmarks, multiple annotation budgets, and ablations of the proposed components.

Critical weaknesses

Applying shared pseudo-label supervision through two detection heads but does not explicitly minimize a discrepancy between their predictions. Its distinction from auxiliary-head supervision could be clarified. Besides, tHow does CCC differ from standard shared-target auxiliary-head supervision, and what evidence supports the claimed cross-consistency mechanism?

Which competing methods were reproduced under the same detector architecture and evaluation protocol?he main comparisons include different detector architectures and results reported by previous studies. The main tables also do not report variability across random seeds, limiting the assessment of relatively small improvements over competing methods.

Questions for authors

How does CCC differ from standard shared-target auxiliary-head supervision, and what evidence supports the claimed cross-consistency mechanism? Which competing methods were reproduced under the same detector architecture and evaluation protocol?

Reviewer_unynOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes IACCL, a teacher–student framework for semi-supervised oriented object detection. It introduces two components. Balanced Pseudo-Labeling (BPL) combines classification confidence with predicted localization quality, estimates class dominance from aggregate confidence mass, and suppresses dominant classes when allocating pseudo-label budgets. Cross Consistency Constraint (CCC) applies the same pseudo-targets to both the student detection head and the frozen EMA-teacher head operating on student features. The auxiliary teacher-head loss updates the student feature extractor without changing the inference architecture. Experiments on DOTA v1.0 and v1.5 report improvements across several labeled-data ratios. Supplementary evaluations examine horizontal detection, an additional detector architecture, and inference efficiency.

Critical strengths

(1) Class imbalance and unreliable pseudo-labels are meaningful challenges in aerial object detection, particularly with limited annotations and highly uneven category frequencies.

(2) BPL clearly defines candidate scores, class dominance, suppression factors, and selection budgets. The supplementary pseudocode makes the selection procedure relatively easy to inspect.

(3) CCC reuses the EMA-teacher head rather than introducing a separately learned auxiliary head. The paper clearly explains that its gradients update student features while teacher-head parameters remain frozen during the auxiliary forward pass.

(4) IACCL achieves the highest reported results in all six partially labeled DOTA settings. Improvements over the strongest listed methods range from 0.76 to 2.29 mAP, suggesting benefits beyond a single annotation budget.

(5) The experiments separately evaluate BPL and CCC, examine composite scoring and the retention scale, and analyze category-level AP and pseudo-label counts. The paper also reports the unresolved failure on the extremely rare container-crane category.

Critical weaknesses

(1) Substituting the definitions into Equation (8) gives

di=min⁡ ⁣{∣Ji∣,⌊βwi∑k∈Jiok⌋}.d_i=\min\!\left\{|J_i|,\left\lfloor \beta w_i\sum_{k\in J_i}o_k \right\rfloor\right\}.

Suppression reduces budgets for dominant classes, while non-dominant classes retain their original budgets. There is no fixed total budget whose unused allocation is explicitly transferred to tail classes. Consequently, increases in tail pseudo-label counts across separately trained models may reflect indirect training dynamics rather than immediate redistribution. The paper should distinguish these effects and compare suppression against a control matched for total pseudo-label count.

(2) Equation (12) adds two pseudo-target losses, while the overall unsupervised weight remains fixed at one. When teacher and student heads are similar, the auxiliary branch can approximately increase the unsupervised gradient applied to the feature extractor, even without a distinct cross-head consistency benefit. The current ablation therefore does not isolate the proposed explanation. Necessary controls include doubling the conventional unsupervised weight, normalizing the two losses by their average, and comparing against a duplicated frozen student head or another matched auxiliary branch.

(3) Table 1 compares many methods using different detectors and potentially different experimental implementations. Although the paper specifies nested labeled splits for its experiments, it does not clearly establish that all compared results use these exact splits and training conditions. The ablation baseline also differs from both Dense Teacher and SOOD in the main table: on DOTA v1.5 at 10% labels, it achieves 50.19 mAP, compared with 46.90 and 48.63, respectively. The manuscript should explain which modifications account for this difference and identify which results were reproduced under a common protocol.

(4) Category-level AP and pseudo-label counts support the proposed motivation, but increased counts alone do not establish improved supervision. The paper should report class-wise pseudo-label precision, recall, localization quality, and retained counts under matched budgets. Because budgets depend on aggregate predicted confidence mass, very rare or poorly predicted classes can still receive few or zero pseudo-labels after flooring. The zero AP for container crane illustrates this limitation. Moreover, assigning zero classification targets to unselected locations may penalize genuine objects excluded by selection; the consequences of this choice should be analyzed.

(5) DOTA v1.0 and v1.5 use the same underlying image split, so they provide less independent evidence than two distinct image collections. Horizontal-box experiments reuse DOTA, and cross-detector gains vary substantially, from 0.96 mAP on Rotated RetinaNet to 3.81 mAP on Rotated FCOS. Independent datasets and repeated labeled splits would strengthen generality claims. Training efficiency also remains unreported: CCC requires an additional head evaluation with gradients propagated into student features. Unchanged inference FLOPs do not establish low training overhead.

Questions for authors

(1) Does BPL explicitly redistribute a fixed pseudo-label budget, or only suppress dominant classes? What happens when total retained pseudo-label counts are matched across variants?

(2) How does CCC compare with doubling the standard unsupervised-loss weight or averaging its two supervision terms?

(3) Does a duplicated frozen student head provide similar gains? How does CCC’s benefit vary with the teacher EMA coefficient and teacher–student divergence?

(4) Which Table 1 results were reproduced using the exact nested splits and training schedule? What distinguishes the ablation baseline from Dense Teacher and SOOD?

(5) Can the authors provide class-wise pseudo-label precision, recall, rotated IoU, and selection counts at common training checkpoints?

(6) How often does flooring produce zero pseudo-label budgets for rare classes? Does ignoring unselected locations outperform assigning zero classification targets?

(7) What is the training-time and memory overhead of CCC, and how stable are results across independent seeds and labeled-data splits?

Reviewer_vG23OpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper provide a student-teacher architecture by using all students as experts to vote, or smoothing the teacher. And they focus on oriented object detection.

The paper argues that dense pseudo-labeling in semi-supervised oriented object detection (SSOOD) inherits the long-tailed class distribution of DOTA, so head classes dominate pseudo-label supervision. It proposes IACCL, which has two parts.

Critical strengths

Relevant problem. Class imbalance in the pseudo-label pool is a real issue for dense teacher-student detectors on long-tailed aerial data, and it is less explored in SSOOD than threshold or consistency design.

Critical weaknesses

I personally think this topic seems a little bit out-dated.

If tried to use DINOV3-Sat to get segmentation + PCA get bounding box, I think it could be a good starting points. And also, at least one can convert Horizontal Box to segmentation mask and using pca to get bounding box.

Another thing is that: I am not experts in this OOD "Oriented Object Detection", I am doing "Out of Distribution" and 3D related problem.

Questions for authors

Is there any specific reason for using resNet50, is there computational resource requirements? Online response requirements?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.