A Multi-correspondence Object Detection Algorithm Based on Keypoints.

Keypoints-based object detection algorithms such as CenterNet and CornerNet have been proved to be very efficient. They use inherent pixels on the feature map to predict whether a pixel is a keypoint of an object, thereby eliminating the need for anchor calculation. However, in cases where the keypoints of multiple objects overlap with each other, Keypoints-based object detection algorithms cannot effectively distinguish them. The reason for the problem is that there is a one-to-one relationship in the network, that is, one location only corresponds to one prediction box. In this paper, we mainly address the problem caused by overlapping center. We achieve our purpose by a kind of multi-correspondence relationship in the network, one location corresponding to multiple prediction boxes. In order to achieve this goal, our work mainly consists of the following two parts. First, we improved the keypoint pooling method. The existing keypoint pooling methods focus on emphasizing the keypoints, but we believe that keypoints detection should suppress the surrounding information while emphasizing the true keypoints. Second, in order to make the corner better match to the correct center, we adopt Pointing Offset to match them. Specifically, the corner will match center by its offset, in order to predict Pointing Offset better, we design a module named corner-to-center (CTC) to improve the matching effect.