Adversarial Alignment for Source Free Object Detection

Source-free object detection (SFOD) aims to transfer a detector pre-trained on label-rich source domain an unlabeled target without seeing data. While most existing SFOD methods generate pseudo labels via source-pretrained model guide training, these usually contain high noises due heavy discrepancy. In order obtain better supervisions, we divide the into source-similar and source-dissimilar parts align them in feature space by adversarial learning.Specifically, design variance-based criterion domain. This is motivated finding that larger variances denote higher recall similarity Then incorporate module mean teacher framework drive spaces of two subsets indistinguishable. Extensive experiments multiple cross-domain datasets demonstrate our proposed method consistently outperforms compared methods. Our implementation available at https://github.com/ChuQiaosong.