acceptodds
Under review as a conference paper at ICLR 2027

ProCreDet: Progressive Credible Sample Selection for Sparsely Annotated Underwater Object Detection

Abstract

Sparsely Annotated Object Detection (SAOD) is designed to address the high cost of dataset annotation. In underwater scenarios, its inherent challenges are particularly pronounced: variable underwater illumination conditions lead to significant discrepancies in instance features across different images, the mixing of unlabeled positive samples into negative sets disrupts the training logic, and the model tends to generate numerous misclassifications and struggles to effectively preserve the inter-class correlation information of targets. These factors collectively impede the improvement of model performance. To tackle these challenges, we propose ProCreDet, a teacher-student framework based on progressive training with credible samples. This framework employs a complementary feature learning mechanism to handle diverse underwater illumination conditions for positive sample mining, integrates a risky sample filtering module to resolve the issue of unlabeled positive samples being mixed into negative samples, and introduces a conflict mitigation module to suppress false predictions while preserving inter-class relationship information, thereby enhancing the generalization ability of the model.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
sparsely annotated object detection, underwater object detection, teacher-student framework, pseudo-labeling, consistency regularization

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_5oMCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ProCreDet, a teacher-student framework for sparsely annotated underwater object detection, where many objects remain unlabeled during training and may incorrectly be treated as background.

The method introduces three components: (1) Risky Sample Filtering (RSF), which dynamically separates reliable positives, uncertain samples, and likely negatives based on teacher confidence distributions; (2) BiMask, which encourages complementary feature learning under different underwater color augmentations; and (3) Conflict Mitigator (CM), which suppresses misleading teacher predictions while retaining useful inter-class information for student supervision.

Experiments on UODD and DUO under different annotation sparsity levels demonstrate improvements over existing methods. With 70% of annotations removed, the YOLO-based implementation achieves 47.2 mAP on UODD and 54.0 mAP on DUO, exceeding the strongest reported competing baselines by 7.6 and 4.9 mAP points, respectively. Ablation studies further support the contributions of the individual modules, particularly RSF, to detection accuracy and training stability.

Critical strengths
  1. Effective handling of missing annotations. The proposed RSF module addresses an important weakness of sparsely annotated object detection: incorrectly treating unlabeled foreground objects as negatives. The interval-purity analysis in Table 6 and the ablation results in Table 7 provide meaningful evidence that filtering uncertain samples improves performance.

  2. Strong experimental improvements under high annotation sparsity. The method achieves substantial improvements on both UODD and DUO, particularly when 70% of annotations are removed. The results across multiple sparsity levels, together with the reported improvements in training stability, demonstrate practical value.

  3. Complementary components with useful ablations. The combination of RSF, BiMask, and CM addresses different challenges in underwater object detection. The ablations demonstrate measurable contributions from the individual components. Furthermore, inference requires only the teacher detector, avoiding additional inference-time branches.

Critical weaknesses
  1. Limited methodological novelty. The framework primarily combines established techniques, including teacher-student learning, confidence-based pseudo-label filtering, feature masking, and consistency regularization. Although the integration is effective, the paper does not sufficiently distinguish its contributions from simpler alternatives. In particular, comparisons against fixed or percentile-based filtering and standard feature dropout would help establish whether the proposed mechanisms provide improvements beyond existing approaches.

  2. Insufficient justification of the statistical assumptions and loss formulation. RSF estimates Gaussian confidence statistics independently for each batch while jointly modeling all categories (Section 3.2). Given the small batch size of two images and the bounded nature of confidence scores, the reliability and calibration of these thresholds deserve further investigation. Moreover, the K-S test uses parameters estimated from the tested samples, making the reported asymptotic p-values potentially inappropriate without correction. Separately, Equations (7)-(10) describe KL divergence directly on logits without clearly specifying normalization and the order of masking and loss reduction. These ambiguities affect reproducibility and the interpretation of the proposed objectives.

  3. Limited evidence of generalization and comparison fairness. The experiments use only two underwater datasets with artificially removed annotations, leaving performance under naturally incomplete annotations unverified. Comparisons also involve different detector architectures and training configurations, making it difficult to isolate architectural advantages from those of the proposed method. Although matched-backbone ablations are provided, stronger controlled comparisons and quantitative cross-domain evaluation would better substantiate the generalization claims.

Questions for authors

Can the authors compare RSF against simpler alternatives, such as fixed confidence thresholds, percentile-based filtering, or moving-average thresholds, while keeping the negative sampling ratio identical? This would help establish the specific contribution of the proposed Gaussian-based selection.

  1. How stable are the estimated Gaussian parameters across batches and categories, particularly with batch size two? Additionally, how are the K-S p-values calculated when the Gaussian parameters are estimated from the same observations?

  2. Could the authors clarify the implementation of the KL divergence in Equations (7)-(10)? Specifically, are the logits normalized before computing KL divergence, and are the sample masks applied before or after loss reduction?

  3. Can the authors provide additional matched-architecture comparisons and evaluate the approach under naturally occurring annotation sparsity or quantitative cross-dataset transfer? Such evidence would strengthen the claims concerning robustness and generalization.

Reviewer_8anCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ProCreDet, a teacher–student framework for sparsely annotated underwater object detection. It combines Risky Sample Filtering (RSF), which uses labeled-instance confidence statistics to identify uncertain unlabeled samples; complementary feature masking (BiMask) across two color-augmented student views; and a Conflict Mitigator (CM), which masks misleading teacher category predictions on labeled instances. Experiments on UODD and DUO under simulated annotation removal show improved detection accuracy, particularly at high sparsity, and reduced post-peak performance degradation. The paper also provides module ablations, hyperparameter analyses, and limited multi-seed results.

Critical strengths
  1. The paper addresses a practically relevant problem. Incomplete annotations can incorrectly turn genuine objects into background training examples. Addressing this issue is useful for underwater detection, where dense objects and poor visibility make exhaustive annotation difficult.

  2. The reported improvements are substantial under high annotation sparsity.

  3. The paper provides several complementary analyses. Module ablations, confidence-interval purity measurements, hyperparameter sensitivity studies, and qualitative comparisons help explain the observed improvements beyond aggregate benchmark scores.

Critical weaknesses
  1. The methodological advance is insufficiently isolated. The framework combines teacher–student learning, confidence-based filtering, complementary masking, and selective distillation. These are understandable design choices, but the experiments do not clearly establish which proposed mechanism provides an advantage over simpler alternatives.

  2. The contribution of RSF is not adequately isolated. RSF combines a dynamic confidence boundary with retaining only 1% of low-confidence negatives. Without comparisons against random subsampling and fixed ignore thresholds at comparable sample counts, it is unclear whether the gains come from the proposed boundary or simply reduced negative supervision.

  3. The treatment of missing annotations in the detection loss is unclear. Equation (10) explicitly filters the consistency loss, but the weak-view branch also uses a conventional detection loss. If uncertain regions remain background targets in that branch, the erroneous supervision motivating the method may persist.

  4. The CM loss is ambiguously defined. The paper describes KL divergence directly on logits without specifying the probability transformation. Moreover, setting excluded logits to zero does not necessarily remove their contribution after normalization. These details are essential for assessing whether CM actually prevents misleading distillation.

  5. The evaluation does not establish effectiveness under realistic missing annotations. Experiments use random annotation removal, whereas real omissions may disproportionately affect small, occluded, or poorly visible objects. Differences in detectors and training settings across methods also make it difficult to isolate the framework’s advantage.

  6. The evidence for progressive sample selection is insufficient. Changing interval proportions does not demonstrate that individual uncertain samples become reliable positives. Furthermore, the risky interval still contains 31.4% of samples late in training, requiring reconciliation with the appendix’s claim that its sample count approaches zero.

Questions for authors
  1. How does RSF compare with random negative subsampling alone and a fixed ignore threshold when the number of retained negatives is comparable?

  2. Are RSF masks applied to the weak-view detection classification loss? Please explain how uncertain regions are treated in both detection and consistency supervision.

  3. What exact probability transformation and masking procedure are used for the KL loss? Are excluded categories removed before normalization or masked in the unreduced loss?

  4. Do the gains persist under size-, occlusion-, or visibility-dependent annotation removal, and under a comparison using the same detector and training schedule?

  5. Does the risky buffer retain samples across iterations or only mask current predictions? What evidence supports progressive sample transitions, and how is the late-stage proportion of 31.4% reconciled with the claim that the sample count approaches zero?

Reviewer_cuaAOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ProCreDet is a student teacher based SAOD framework for underwater object detection. SAOD is the problem of partially labeled datasets where each image can have missing positives leading to ambiguity between real positive and negative regions in an image which is important while training object detectors that require dense losses. ProCreDet solves this by proposing sample filtering, augmentations and masking and multiple views of the image for consistency. The authors also propose conflict mitigator to filter loss on annotated positives. Their experiments on UODD and DUO with RetinaNet and YOLOv8 report higher accuracies.

Critical strengths
  1. The paper proposes a reasonable strategy to mitigate the missing positive issue in SAOD where unlabeled regions can get treated incorrectly. They propose using score based filtering, feature masking and consistency loss.
  2. At high sparsity settings like 70% ProCreDet/RetinaNet reaches 43.2 mAP on UODD versus 36.2 for BRL, and 50.0 on DUO versus 45.4 for plain RetinaNet.
  3. The authors also provide good ablations and experiments to validate that the approach shows improvements over the two datasets for a few different architectures.
Critical weaknesses
  1. In Table 3, a bunch of baselines seem to have lower performance than the vanilla model trained on the sparse data. I feel this might be a tuning issue then and make the comparisons weak. BRL and Co-mining trail vanilla RetinaNet at every sparsity (e.g., 44.8 and 40.2 vs 45.4 at 70%), and Co-Student and S²Teacher trail vanilla FCOS at 70% (41.0 and 45.5 vs 47.9).
  2. Missing results on standard datasets like COCO and VOC which most baselines have previously reported. Not reporting makes it questionable if the improvements are specific to the new datasets being used.
  3. Checkpoint selection is unclear, as no val split is defined but "best training validation point" is used. It would be good to clarify how this is measured.
  4. L323 mentions 5 detectors but I can only find 2 (RetinaNet and YOLOv8). Table 4 only lists fully supervised oracles.
  5. The authors claim prior works overlook false negatives but Co-mining, BRL and SparseDet take those into account and even mitigate the negative effects.
  6. RSF filters the consistency loss (Eq. 10), but it is unclear whether risky unlabeled objects are still treated as background in the detection loss (Eq. 11).

Minor - Typos in Figures 2 and 7 ("Labeld", "ProCredDet"); "7.6%" should be "7.6 points" (L87)

Questions for authors

I would like to request the authors to address the questions I raised in weakness around baseline tuning, standard benchmarks for fair comparison, clarifications around a few things mentioned in the paper and proper discussions with the baselines.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.