A Free Lunch for Unsupervised Domain Adaptive Object Detection without Source Data

Unsupervised domain adaptation (UDA) assumes that source and target data are freely available usually trained together to reduce the gap. However, considering privacy inefficiency of transmission, it is impractical in real scenarios. Hence, draws our eyes optimize network without accessing labeled data. To explore this direction object detection, for first time, we propose a data-free adaptive detection (SFOD) framework via modeling into problem learning with noisy labels. Generally, straightforward method leverage pre-trained from generate pseudo labels optimization. difficult evaluate quality since no domain. In paper, self-entropy descent (SED) metric proposed search an appropriate confidence threshold reliable label generation using any handcrafted Nonetheless, completely clean still unattainable. After thorough experimental analysis, false negatives found dominate generated Undoubtedly, mining helpful performance improvement, ease simulation through augmentation like Mosaic. Extensive experiments conducted four representative tasks have demonstrated can easily achieve state-of-the-art performance. From another view, also reminds UDA community not fully exploited existing methods.