Decompose to Adapt: Cross-Domain Object Detection Via Feature Disentanglement

Recent advances in unsupervised domain adaptation (UDA) techniques have witnessed great success cross-domain computer vision tasks, enhancing the generalization ability of data-driven deep learning architectures by bridging distribution gaps. For UDA-based object detection methods, majority them alleviate bias inducing domain-invariant feature generation via adversarial strategy. However, their discriminators limited classification due to unstable training process. Therefore, extracted features induced cannot be perfectly and still contain domain-private factors, bringing obstacles further discrepancy. To tackle this issue, we design a Domain Disentanglement Faster-RCNN (DDF) eliminate source-specific information for task learning. Our DDF method facilitates disentanglement at global local stages, with Global Triplet (GTD) module an Instance Similarity (ISD) module, respectively. By outperforming state-of-the-art methods on four benchmark UDA our is demonstrated effective wide applicability.