A Modified Loss Function Approach for Instance Segmentation Improvement and Application in Fish Markets.

This paper presents an approach to image segmentation and classification algorithm where the dataset has only few images labelled, done intentionally. The method tries to classify only the few instances with enough quality in the image, the KeyFish. In order to not being punished with wrong false positives, it must learn the examples but not the context. The application is intended for wholesale fish markets. Due to the depth and occlusions of the fish tray, the camera can only visualize a small fraction of the total instances. The main goal is to predict in the best possible quality the few fish seen, regardless of other occurrences. Tests have been made over Yolact++ architecture and the proposed method, with an increase in precision from 85.97% to 89.69% in bounding box and 74.03% to 81.42% in detection of the mask for a 50% overlap limit.