Boundary Information Aggregation and Adaptive Keypoint Combination Enhanced Object Detection

• Published in: Image and Graphics Vol. 12888; pp. 148 - 159
• Main Authors: Zhao, Ping, Yao, Dongsheng, Sun, Lijun, Fan, Jiaqi, Chen, Panyue, Wei, Zhihua
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2021; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Keypoint prediction; Object detection
• ISBN: 3030873544; 9783030873547
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-87355-4_13

Keypoint-based methods achieve increasing attention and competitive performance in the field of object detection. In this paper, we propose a new keypoint-based object detection method in order to better locate center keypoints of objects and adaptively combine keypoints to obtain more accurate bounding boxes. Specifically, to better locate center keypoints of objects, we aggregate boundary information by adding the center pooling operation to the original center keypoints prediction branch. The boundary information is the location of object boundary which is more easier to predict than object center. Furthermore, to obtain more accurate bounding boxes, we propose an adaptive keypoint combination algorithm to map all keypoints back to the original image so that the keypoints are combined with less localization errors. Experiments have demonstrated the effectiveness of the our proposed methods.