Soft focal loss: Evaluating sample quality for dense object detection

Abstract

Classic detectors divide the candidate boxes into positive and negative groups based on their intersection-over-union (IoU) with matched objects. Such a sharp label assignment method does not directly consider the distance between the centers of the two boxes. Consequently, some positive samples are ambiguous, which limits the detection performance. In this paper, we propose a new label assignment method by considering two different perspectives. The first perspective is IoU to indicate the degree of overlap. The other one is the defined variable to characterize the center-distance between the candidate box and its matched ground truth. In addition, the classification is usually optimized by Focal Loss for paying more attention to hard examples, but it affects the training of high-quality samples. Therefore, we define the Soft Focal Loss (SFL) and the quality factor that reflects the quality of the samples. Embedding the quality factor into SFL makes the network focus on learning high-quality rather than hard examples. Furthermore, the quality factor is utilized to re-weight the classification and regression losses to enhance the correlation between these two tasks. Experiments on COCO show that the proposed approach can improve RetinaNet by 1.3% and 1.2% AP with backbone ResNet-50 and ResNet-101 in 1x training schedule, without incurring any additional overhead.