Delving Into High-Quality Domain-Adaptive Object Detection With Denoised Hierarchical Learning

Domain adaptive object detection (DAOD) methods have demonstrated promising performance in detecting objects in unseen scenarios by transferring knowledge from a labeled source domain to an unlabeled target domain. Recent graph matching-based DAOD methods extract instance-level features and perform graph-based interaction and alignment for domain adaptation. However, their effectiveness remains constrained by hierarchical mismatches and noisy samples throughout the iterative processes of feature completion, memory bank update, and category-level adaptation. To address these limitations, we propose Denoised Hierarchical Learning (DHL) to achieve high-quality domain-adaptive object detection. Specifically, to avoid hierarchical mismatch, hierarchical feature completion (HFC) generates hallucinated features for missing categories by leveraging the same hierarchical features and memory bank, thus ensuring identical category participation during adaptation. Then, hallucinated feature denoising (HFD) leverages a denoised classification network and an outlier detection mechanism to eliminate noise mixed into the hallucinated features. To enhance the memory bank representation, the hierarchical denoised memory bank (HDMB) leverages the same-hierarchy instance-level features and the denoised classification network to fully utilize all features and denoise the update process. Furthermore, to enhance category-level adaptation in graph matching, prior-agnostic semantic-aware graph matching projects hierarchical and domain-specific features into a unified feature space and learns category-level representations via supervised contrastive loss. Extensive experiments on eight adaptation tasks show that DHL provides its largest gains under challenging nighttime shifts while achieving competitive performance under weather, location, and style shifts with substantially reduced training time. Source code and configuration files are available at https://github.com/GingerCohle/DHL