Knowledge-Driven Network for Object Detection

Object detection is a challenging computer vision task with numerous real-world applications. In recent years, the concept of object relationship model has become helpful for and been verified realized in deep learning. Nonetheless, most approaches to modeling relations are limited using anchor-based algorithms; they cannot be directly migrated anchor-free frameworks. The reason that algorithms used eliminate complex design anchors predict heatmaps represent locations keypoints different categories, without considering between keypoints. Therefore, better fuse information heatmap channels, it important visual this paper, we present knowledge-driven network (KDNet)—a new architecture can aggregate keypoint augment features detection. Specifically, processes set simultaneously through interactions their local geometric features, thereby allowing relationship. Finally, updated were obtain corners objects determine positions. experimental results conducted on RIDER dataset confirm effectiveness proposed KDNet, which significantly outperformed other state-of-the-art methods.