Enhanced RT -DERT Automatic annotation for Under water Images

Manual annotation of fish in underwater video is time-consuming and becomes difficult when frames contain small fish, low illumination, complex backgrounds, and coral structures. The difficulty is further increased by underwater-specific visual conditions such as color distortion, reduced visibility, background clutter, occlusion, and variations in the appearance of the same object across different underwater environments. while examining the limitations of detector generalization across different underwater domains. The study follows a progressive experimental methodology. Initially, conventional deep-learning object detectors were investigated to establish a baseline for underwater fish detection. YOLO11n, Faster R-CNN, RetinaNet, and SSD300 were evaluated under the experimental dataset conditions, providing a comparison of detection performance and computational characteristics. Based on the baseline investigation, the study was extended toward an end-to-end transformer-based detection framework using RT-DETR with a PResNet-18 backbone and a single fish class. The strongest examined RT-DETR checkpoint was obtained at epoch 50, achieving a COCO AP@[0.50:0.95] of 0.28847 and an AP50 of 0.60264. The trained detector was subsequently employed as an automatic annotation system for underwater video frames, where confidence-based filtering and duplicate prediction handling were applied to produce fish bounding boxes. An initial automatic annotation experiment involving 61 frames generated 466 predicted fish bounding boxes, demonstrating the practical feasibility of using the trained detector as an annotation-assistance tool. A longer real underwater video containing 2,668 frames was also processed, resulting in 2,247 frames with detected fish and 4,984 predicted fish bounding boxes. These detection counts demonstrate the operational use of the proposed annotation workflow, although they are not considered direct measures of detection accuracy because complete ground-truth annotations were not available for the full video sequences. To investigate real-world generalization, the trained detector was further evaluated on a separate underwater video with different visual characteristics. The experiment revealed significant domain-shift effects, including missed fish, failure to detect small fish, and false detections on coral and other background structures. These observations indicate that changing the confidence threshold alone is insufficient to address the problem because reducing the threshold can increase the number of background false positives. In response to this limitation, a target-domain refinement dataset was developed through manual annotation using Label Studio. Twenty-five target-domain frames were successfully annotated with 253 fish instances using a single fish class while explicitly excluding coral and background structures. The verified dataset was further divided into a 20-image training subset containing 91 fish annotations and a five-image unseen validation subset containing 162 fish annotations. The overall findings demonstrate that deep-learning-based automatic fish annotation is feasible for underwater video processing, but reliable cross-domain deployment remains strongly dependent on the similarity between training and target environments. The study therefore highlights domain shift as a major practical challenge in underwater object detection and establishes a target-domain annotation and refinement workflow for addressing this limitation. The present work focuses on the integration and evaluation of established detection methods rather than proposing a new detector architecture.