RSOD: Reliability-Guided Sonar Image Object Detection with Extremely Limited Labels

AI summary

Overview

Research area: Computer vision, specifically semi-supervised object detection (SSOD) applied to underwater forward-looking sonar imagery.

Technical level: Advanced. The paper assumes familiarity with teacher-student semi-supervised detection, pseudo-labeling, consistency regularization, Faster R-CNN with FPN, and detection loss design.

Scope: The paper proposes RSOD, a reliability-guided teacher-student framework for sonar object detection under extremely limited annotation budgets, and introduces a new sonar dataset called FSOD.

What This Paper Is About

Sonar images are hard to annotate because they have few texture details and are noisy, so non-experts cannot reliably distinguish subtle differences between object classes. This makes it difficult for supervised detectors to get enough labeled data, and it also causes the pseudo-labels that semi-supervised methods generate to be noisy. The paper's goal is to build a semi-supervised detection framework that learns from unlabeled sonar images well enough to perform competitively when only a tiny fraction of the data is annotated.

Key Contributions

1. The authors state they are the first to introduce a semi-supervised object detection algorithm specifically designed for sonar images, with an RSOD framework that uses unlabeled samples to lower annotation costs while improving detector performance.

2. They propose a pseudo-label reliability score computed from the teacher's prediction consistency across multiple augmented views, together with an object mixed pseudo-label (OMPL) strategy intended to maximize the use of high-quality pseudo-labels and minimize the effect of noisy ones on the loss.

3. They release a new dataset, the Forward-Looking Sonar Image Object Detection (FSOD) dataset, containing 3,929 sonar images captured with an Oculus M750d sonar and annotated in 10 object categories.

4. They report experiments across multiple labeled-data ratios on the UATD and FSOD datasets showing RSOD outperforms existing state-of-the-art methods, with what they describe as significant improvements when only 5% of the data is annotated.

Main Findings

• Headline UATD claim: On the UATD dataset, RSOD trained with only 5% labeled data achieves results the authors describe as able to compete against the baseline algorithm trained on 100% labeled data. The specific 100%-label baseline number is not reported in the provided content.

• UATD mAP at 1%, 5%, 10% labels: RSOD reaches 23.2, 34.7, and 34.8 mAP. The fully supervised Faster R-CNN baseline gets 13.7, 22.9, and 26.2. Other SSOD baselines are lower than RSOD at each setting: Soft Teacher 23.0/30.8/33.1; Unbiased Teacher V2 21.6/31.0/33.4; Pseco 22.9/32.7/33.6; MixPL 22.2/31.6/32.2; VC 21.5/30.9/32.2.

• FSOD mAP at 5%, 10%, 20% labels: RSOD reaches 67.4, 72.6, and 75.5. Faster R-CNN gets 44.9, 50.6, and 59.6; Soft Teacher 56.4/65.7/72.2; Unbiased Teacher V2 65.0/70.1/72.8; Pseco 65.7/70.5/74.6; MixPL 64.3/70.2/74.5; VC 65.9/69.6/74.0.

• Small-object detection on UATD (mAP(small)): RSOD scores 10.1, 12.6, and 12.8 at 1%, 5%, and 10% labels, against 2.9, 7.1, and 7.3 for Faster R-CNN. Best competing baselines at 1% reach 9.9 (Soft Teacher), at 5% 12.2 (VC), and at 10% 12.6 (Unbiased Teacher V2).

• Ablation on FSOD with 5% labels: The fully supervised Faster R-CNN baseline (M1) scores 44.9 mAP. Adding the teacher-student network (M2, EMA) raises this to 64.3. Adding PLRW (M3) gives 65.2, adding CPRL (M4) gives 66.3, and adding OMPL (M5) yields the final 67.4.

• Chosen hyperparameters: The reliability scaling factor H is set to 0.6, the reliability score threshold to 0.5, and the IoU threshold to 0.2, which produces the best result of 67.4 on FSOD with 5% labels. The number of augmented views is 2 (mirrored and scaled).

• Qualitative observations: Soft Teacher and VC are reported to suffer missed detections on small sonar objects, while Pseco and MixPL produce false detections with high confidence. RSOD is reported to avoid both, matching ground truth.

• New dataset statistics: FSOD contains 3,929 images across 10 categories with a long-tailed distribution. The steel frame category has the most samples at 1,231 images (31.5%), while the diver category has the fewest at 81 images (2.1%). Unlike UATD's mannequin samples, FSOD includes a diver category.

Methodology in Plain English

The approach follows a two-network teacher-student pattern. A teacher network looks at unlabeled sonar images and proposes bounding boxes with class confidences. Those proposals become pseudo-labels used to train a student network, together with the small set of genuinely labeled images. The total training objective combines a supervised loss on labeled data with an unsupervised loss on pseudo-labels, weighted by a hyperparameter.

The paper's first idea is to stop trusting pseudo-labels just because the teacher is confident. Instead, the teacher is run on the original image and on augmented views (mirrored and scaled). Each predicted box is matched to its best counterpart in the augmented views by IoU, and a reliability score combines how well the boxes overlap, how close their class confidences are, and whether their predicted classes agree. The result is passed through a Sigmoid and scaled by a factor H. Boxes whose predictions hold up across views score higher.

The second idea is a mixing trick. High-reliability boxes from several images in a batch are rotated, scaled, and pasted onto the first image of the batch, with a check to avoid overlapping existing boxes. This synthesizes scenes with more small and multiple objects, which the authors argue helps the student learn challenging cases that pseudo-labels alone under-serve.

The third idea is how reliability feeds the loss. Each pseudo-label contributes to the unsupervised loss in proportion to its reliability score. In addition, for boxes above the reliability threshold, a corner-point regression loss directly aligns the corners of the student's predicted boxes with the teacher's, normalized by the diagonal of the minimum enclosing rectangle of the matched pair.

Training uses Faster R-CNN with FPN and a ResNet-50 backbone initialized from ImageNet, following the STAC protocol. Asymmetric augmentation is applied: weak augmentation (random flipping) for the teacher's unlabeled inputs and strong augmentation (greater degree of random flipping plus random Gaussian blur) for the student. Models train for 180K iterations on a single NVIDIA RTX 4090 GPU using SGD with learning rate 0.001, momentum 0.9, and weight decay 0.0001. Evaluation uses standard mean average precision (mAP) on UATD and FSOD.

Data collection for FSOD took place in Bohai Bay using a remotely operated vehicle (ROV) equipped with an Oculus M750d multibeam forward-looking sonar. Objects were connected and suspended by ropes at depths of approximately 3 to 20 meters below the surface and placed at fixed intervals, imaged from different angles and distances of 2 to 15 meters. UATD used its original partition: 7,600 training images and 800 images each in the test and validation sets. FSOD was split 6:2:2 into training, validation, and test sets.

Why This Matters

Impact on research. The paper argues that standard semi-supervised detection methods were not designed for the texture-poor, noise-heavy character of sonar imagery, where confidence thresholds alone admit coarse or outright wrong boxes. By re-scoring pseudo-labels through multi-view consistency and by re-weighting losses accordingly, the work suggests a route to high accuracy under annotation budgets as low as 1% and 5%. The release of FSOD, including a diver category not present in UATD's mannequin samples, adds a public resource to a field the authors describe as hampered by scarce public data.

Real-world applications:

• Underwater accident rescue and emergency response, where detecting divers quickly matters.
• Equipment maintenance and inspection of submerged infrastructure such as steel frames.
• Biological surveys of underwater environments.
• Diving safety assurance and accident prevention systems.

Industry relevance. Sonar data is described as highly confidential and costly to annotate, which limits supervised pipelines in practice. A method that performs well with 5% of labels can substantially cut labeling spend for operators of underwater robots and ROVs. Because the framework builds on a standard Faster R-CNN backbone, it is compatible with existing detection stacks rather than requiring a bespoke architecture.

Future Directions

• How far can the label budget shrink? The lowest tested setting is 1% on UATD; the paper's smallest reported setting on FSOD is 5%. Whether the reliability mechanism holds at even lower ratios is untested.

• Generalization across sonar hardware and sites. FSOD was collected in Bohai Bay with an Oculus M750d. Whether the reliability scores transfer to other sonar models, frequencies, or sea conditions is not established.

• Handling the long tail. FSOD's diver category has only 81 images (2.1%), while steel frame has 1,231 (31.5%). The paper does not report per-class results, so it is unclear how much the method helps the rarest categories.

• Sensitivity to the new hyperparameters. Three parameters were introduced (H, the reliability threshold, and the IoU threshold), tuned to 0.6, 0.5, and 0.2 on FSOD with 5% labels. How these values behave across datasets and label ratios remains an open question.

Target Audience

This paper is most useful to researchers and engineers working on underwater perception, autonomous underwater vehicles, and ROV-based inspection systems, particularly those facing annotation bottlenecks. It also suits semi-supervised object detection researchers looking for a domain where standard confidence-threshold filtering breaks down, and practitioners who want a recipe that starts from an off-the-shelf Faster R-CNN with FPN and ResNet-50 backbone. Readers without a background in detection losses and teacher-student training will find the methodology section demanding.

Authors’ abstract

Object detection in sonar images is a key technology in underwater detection systems. Compared to natural images, sonar images contain fewer texture details and are more susceptible to noise, making it difficult for non-experts to distinguish subtle differences between classes. This leads to their inability to provide precise annotation data for sonar images. Therefore, designing effective object detection methods for sonar images with extremely limited labels is particularly important. To address this, we propose a teacher-student framework called RSOD, which aims to fully learn the characteristics of sonar images and develop a pseudo-label strategy suitable for these images to mitigate the impact of limited labels. First, RSOD calculates a reliability score by assessing the consistency of the teacher's predictions across different views. To leverage this score, we introduce an object mixed pseudo-label method to tackle the shortage of labeled data in sonar images. Finally, we optimize the performance of the student by implementing a reliability-guided adaptive constraint. By taking full advantage of unlabeled data, the student can perform well even in situations with extremely limited labels. Notably, on the UATD dataset, our method, using only 5% of labeled data, achieves results that can compete against those of our baseline algorithm trained on 100% labeled data. We also collected a new dataset to provide more valuable data for research in the field of sonar.