Foreground-aware Universe Graph Matching for Domain Adaptive Object Detection.

Abstract

Domain Adaptive Object Detection (DAOD) focuses on improving the generalization ability of cross-domain object detection via knowledge transfer. Recent advances in DAOD strive to minimize the domain discrepancy for adapting class-conditional distributions by aligning cross-domain sampled node pairs in the non-Euclidean graphical space. However, these methods of reformulating the adaptation with graph matching may fail to concentrate on precise semantics alignment, due to the domain-biased graph modeling and the unreliable matching with background nodes. To solve these issues, a Foreground-aware Universe Graph Matching (FUGM) framework is proposed for DAOD. Specifically, a virtual universe graph is category-wise constructed for modeling semantic knowledge, and node representation is refined via Collaborative Graph Reasoning (CGR), which incorporates with self-loop calibration for the information that node learns about itself and universe graph based semantic interaction. To reduce the spurious matching pair for better semantics alignment, Universe Graph Matching (UGM) is developed to encourage instance nodes to match anchor nodes and remove absorbing nodes. The resulting matching model enables end-to-end learning especially for instance node pairwise affinity enhancement through gathering foreground nodes of corresponding category together. Extensive experiments verify that FUGM outperforms existing works significantly.