Contrastive Learning

This paper presents ContrastZSD, a semantics-guided contrastive network for zero-shot object detection (ZSD). The framework improves visual-semantic alignment and mitigates the bias problem towards seen classes by incorporating region-category and …