AN IMPROVED ATSS ALGORITHM BASED ON CORNER AND CENTER POINTS AND ITS APPLICATION FOR OBJECT DETECTION IN COMPLEX SCENES

Recently, object detection algorithms that rely on convolutional neural networks have gradually replaced conventional object detection approaches, and have become mainstream algorithms. Most object detection algorithms re-quire dense anchors over the image, then the anchors are separated into positive or negative samples to predict the bounding box position offset as well as the classification confidence. Dense anchors greatly improve the recall rate of models, but bring some problems at the same time, such as imbalance in quantity between positive and negative samples, which reduce the efficiency of network and accuracy. In addition, positive samples of high quality are more beneficial to the bounding box regression, but the factor is often ignored. In order to further alleviate this imbalance and select anchors with higher quality as positive samples, this paper proposes Improved Adaptive Training Sample Selection with Corner and Center Points (CC-IATSS). CC-IATSS takes the shape of the anchor into account by introducing the distance of the corner point. We first demonstrate that the shape and position of anchors are crucial factors, which directly reflect the quality of positive sample. Then, we calculate the distance of the center and corner points between anchors and the Ground Truth (GT) bounding boxes. Finally, we combine both results to get the anchors close to the GT bounding box in shape and position. Experimental results show that CC-IATSS achieves around 1.4% mAP improvement on PASCAL VOC and 0.9% on MS COCO compared to ATSS.