Description
RetinaMask is a one-stage object detection method that improves upon RetinaNet by adding the task of instance mask prediction during training, as well as an adaptive loss that improves robustness to parameter choice during training, and including more difficult examples in training.