Zero-shot object detection with contrastive semantic association network

Zero-shot object detection (ZSD) is dedicated to the task of precisely localizing and identifying unfamiliar objects that have not been encountered before. In this paper, a contrastive semantic association network is proposed to address the knowledge transfer challenge from seen classes to unseen ones in ZSD. It enables efficient information propagation through similarity-based connections, thereby establishing a clearer link between seen and unseen categories. Moreover, a visual-semantic contrastive learning technique is developed to mitigate the node convergence issue caused by the graph structure of the proposed network. By emphasizing the visual and semantic distinctiveness across different categories, the proposed model leverages semantic information and graph structure knowledge to enhance the generalization capability of seen and unseen feature projection. Extensive experiments demonstrate the superior performance of our model compared to other zero-shot object detection methods, showcasing notable improvement in mean average precision (mAP) on the MS-COCO dataset. The code and models are publicly available at: https://github.com/lihh1023/CSA-ZSD/tree/master.