Class-Aware Memory Guided Unbiased Weighting for Universal Domain Adaptive Object Detection
Cross-domain object detection aims to align the feature distributions across the source and target domains. Existing cross-domain object detectors typically rely on identical label space assumption, which, however, greatly limits their universality under class gap. This paper introduces Universal Domain Adaptive Object Detection (UDAOD) toward more practical scenarios without any prior knowledge on the category consistency. In the proposed universal setting, the category space is partially intersected (i.e., common classes) between domains. The class gap caused by source-private and target-private classes leads to serious negative transfer and degrades adaptation performance. To this end, we propose a Universal Cross-domain Faster RCNN (UCF) with a novel unbiased weighting mechanism to effectively measure the common or private classes. Specifically, we propose a dynamic Class-aware Memory (CaM) to overcome the bias of class weights, caused by class incompleteness in a batch of UniDA. We further propose a Weight Surgery Equalization (WSE) to strengthen the polarization of the weights for common and private classes and suppress incorrect alignment. Extensive experiments under the novel UDAOD setting on multiple benchmarks including PASCAL VOC, Clipart, WaterColor, Cityscapes, and FoggyCityscapes are implemented, which shows the SOTA universality of our model.