DropLoss for Long-Tail Instance Segmentation

Long-tailed class distributions are prevalent among the practical applications of object detection and instance segmentation. Prior work in long-tail segmentation addresses imbalance losses between rare frequent categories by reducing penalty for a model incorrectly predicting label. We demonstrate that heavily suppressed correct background predictions, which reduces probability all foreground with equal weight. Due to relative infrequency categories, this leads an biases towards more categories. Based on insight, we develop DropLoss -- novel adaptive loss compensate without trade-off With loss, show state-of-the-art mAP across rare, common, LVIS dataset. Codes available at https://github.com/timy90022/DropLoss.