CD-Buffer: Complementary Dual-Buffer Framework for Test-Time Adaptation in Adverse Weather Object Detection

Test-Time Adaptation (TTA) enables real-time adaptation to domain shifts without off-line retraining. Recent TTA methods have predominantly explored additive approaches that introduce lightweight modules for feature refinement, while subtractive approaches remove domain-sensitive channels as an alternative direction. We observe that these paradigms exhibit complementary effectiveness patterns: subtractive methods excel under severe shifts by removing corrupted features, while additive methods are effective under moderate shifts requiring refinement. However, each paradigm operates effectively only within limited shift severity ranges, failing to generalize across diverse corruption levels.

We propose CD-Buffer, a complementary dual-buffer framework where subtractive and additive mechanisms operate in opposite yet coordinated directions driven by a unified discrepancy metric. Our framework couples removal and refinement through discrepancy-driven channel-wise balancing, automatically adapting to heterogeneous feature-level domain shifts without manual tuning. Extensive experiments on KITTI, Cityscapes, and ACDC demonstrate state-of-the-art performance across diverse adverse weather conditions and severity levels.

CD-Buffer results on adverse weather conditions. Object detection results are visualized with bounding boxes under diverse weather conditions. We compare ground truth (GT), Direct Test (source model without adaptation), and our CD-Buffer results, demonstrating effective adaptation across varying domain shift severities.

Complementary Dual-Buffer Framework. CD-Buffer computes feature-level domain discrepancy by comparing target feature statistics with pre-computed source statistics. This discrepancy drives two coordinated adaptation modules. The Subtractive Buffer suppresses severely shifted channels via discrepancy-weighted regularization on learnable mask scores, while the Additive Buffer provides adaptive compensation through lightweight convolutions with inverse soft masking.

This unified design enables differentiated channel-wise treatment: severely shifted channels are suppressed and compensated, moderately shifted channels are refined, and stable channels receive minimal intervention. The result is a robust TTA framework that consistently handles heterogeneous domain shift magnitudes in adverse weather object detection.