A reproduction of the detector training in Bodine et al. (2026), “GhostVision: Democratizing Derelict Gear Detection Using Low-Cost Sonar and Artificial Intelligence,” as a baseline for my OCE495 capstone team. The method, dataset, and settings are theirs. The code is four scripts (convert, train, evaluate, visualize), written mostly with Claude. Neither YOLO model matched the paper’s test scores, and my YOLOv26s scored well above theirs (mAP@50 of 0.40 against 0.07).

YOLOv26s detections on twelve test images YOLOv26s on the test split: clean hits (top), worst misses (middle), worst false alarms (bottom). Sonar images from the Bodine et al. dataset, CC BY-SA 4.0; this figure carries the same license.

Precision, recall, and F1 versus confidence threshold Precision, recall, and F1 against confidence cutoff. YOLOv12s peaks at 0.04, so the 0.25 default throws away many of its hits.

Training and validation loss and validation mAP per epoch Training history. Validation loss flattens within about 10 epochs.