THNet: Transferability-Aware Hierarchical Network for Robust Cross-Domain Object Detection

Abstract

Deep learning has advanced object detection, but generalizing models from source to target domains remains a challenge due to multi-level domain drift and untransferable information. To address this, we propose a transferability-aware hierarchical domain-consistent object detector (THNet), incorporating instance-level, pixel-level, and image-level alignment subnets for robust cross-domain detection. THNet first aligns local foreground-transferable features through pixel-level adversarial learning and foreground-aware attention, then captures global domain-invariant features via image-level subnet with channel-transferable attention. Additionally, a prototype graph convolutional network alleviates instance distribution differences by maximizing inter-class distances and minimizing intra-class distances. A domain-consistent loss harmonizes training for better convergence in multi-level domain alignment. Extensive experiments demonstrate that THNet outperforms state-of-the-art methods on multiple cross-domain datasets, achieving top accuracies of 51.9%, 46.0%, 41.2%, and 51.9% across different tasks.

Keywords

• Cross-domain object detection
• hierarchical domain alignment
• domain-consistent loss
• transferable attention
• adversarial learning