Fish Detection and Size Estimation Using Curved Ellipse Modeling and Voting-Based Matching

The paper presents a new approach to reduce bycatch in the fishing industry by automating the catch process through species recognition and size estimation. A segmentation model is trained on our own dataset to detect a selection of fish species using the state-of-the-art RTMDet model as a base to achieve APseg of 0.762 and APbb of 0.755. A novel curved-ellipse model is developed that allows appropriate keypoint detection beyond the field of view, a limitation not addressed in previous works. A novel voting-based algorithm is developed to robustly pair instances when multiple detections are made and reject low-confidence matches. Finally, standard stereo projection of underwater cameras are applied with taking into account the degree of bending of the fish to estimate the true length. The fish species, confidence level, and length can then be passed to the decision-making component for discriminative bycatch reduction. End-to-end inference speed per image pair has been tested at ≈0.12s on workstation and ≈0.35s on GPU edge device.