Domain Contrast for Domain Adaptive Object Detection

Despite of the substantial progress visual object detection, models trained in one video domain often fail to generalize well others due change camera configurations, lighting conditions, and person views. In this paper, we present Domain Contrast (DC), a simple yet effective approach inspired by contrastive learning for training adaptive detectors. DC is deduced from error bound minimization perspective transferred model, implemented with cross-domain contrast loss which plug-and-play. By minimizing loss, transfers detectors across domains while naturally alleviating class imbalance issue target domain. can be applied at either image level or region level, consistently improving detectors’ discriminability maintaining transferability. Extensive experiments on commonly used benchmarks show that improves baseline state-of-the-art significant margins, demonstrating great potential large divergence. Code released https://github.com/PhoneSix/Domain-Contrast .