Domain-Aware Feature Distillation for Cross-Domain Object Detection

Abstract

Self-training with teacher-generated pseudo-labels has driven notable progress in cross-domain object detection. However, most existing methods primarily rely on output-level supervision, while overlooking the transfer of intermediate representations. Although feature distillation appears to be a natural solution, directly applying conventional distillation under domain shifts often leads to limited gains. In this paper, we identify two key challenges that hinder effective distillation: (1) deceptive channel noise caused by domain discrepancy, and (2) blind spatial alignment in the absence of target annotations. To address these issues, we propose Domain-Aware Feature Distillation (DAFD). Specifically, DAFD consists of Dual-Perspective Channel Refinement (DPCR), which jointly evaluates channel structural quality and student-teacher spatial agreement to suppress unreliable channel responses. Concurrently, Collaborative Confidence Region Mining (CCRM), which dynamically mines discriminative spatial masks to focus distillation on informative regions. Through domain-aware channel purification and spatial calibration, DAFD enables more reliable feature transfer across domains. Extensive experiments on standard cross-domain benchmarks demonstrate that DAFD achieves state-of-the-art adaptation performance across diverse detectors. We further verify its effectiveness in the more challenging source-free object detection.