An anchor-free detector designed for dense small water object detection

Abstract

The detection of dense small objects on the water surface is one of the hot topics in object detection. In this paper, a one-to-many label assignment strategy based on the OTA algorithm, which is applied to anchor-free detector, is proposed to improve the detection accuracy of dense small objects on water surface. To be specific, one-to-many label assignment means that a ground truth (GT) corresponding to multiple prediction boxes is conducted globally. The cost function used by OTA algorithm is improved to make the distribution of positive and negative samples more reasonable. Meanwhile, an efficient training strategy is designed to accelerate network convergence by adding L1 loss in the last stage of training. The results show that the proposed strategies achieve 87.9% average precision (AP). Compared to the original algorithm, we achieve 3.3% relative improvement for the detection precision of dense small objects.