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

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