DCRT: Depth-confidence-guided complementary residual temporal adapter for camera-based BEV 3D object detection

Abstract Camera-based bird's-eye-view (BEV) 3D object detection is important for autonomous-driving perception, but the quality of BEV representations is easily affected by depth-estimation errors, object motion, and cross-frame alignment errors. Existing temporal BEV detection methods usually introduce historical frames to enhance the current representation. However, historical BEV features do not always provide effective complementary value and may introduce misaligned or redundant responses under occlusion changes, dynamic object displacement, and local depth uncertainty. To address this issue, this paper proposes a Depth-Confidence-Guided Complementary Residual Temporal Adapter (DCRT). DCRT is integrated into a baseline temporal BEV detection framework as a bypass residual module. It uses depth-confidence cues to construct complementary relationships between the current and historical frames, and generates historical residual candidates through dynamic local sampling, enabling selective compensation using historical information that provides effective complementary value. In addition, DCRT introduces a Null-aware residual arbitration mechanism to suppress low-quality residual injection when historical information lacks effective complementary value. Experiments on the nuScenes and Waymo validation sets show that DCRT consistently improves the BEVDepth + T temporal baseline under their respective evaluation protocols. Ablation studies, complexity analysis, and visualization results further validate the effectiveness of the proposed method.