Efficient Deformable Temporal Alignment for Bird’s-Eye-View 3D Object Detection
Year: 2026
Type: conference-paper
Abstract: Temporal fusion is central to Bird’s-eye-view (BEV) 3D object detection, yet it often produces ghosting artifacts and gradual misalignment over long horizons when it relies on rigid ego-motion compensation and a fixed sampling radius. These assumptions break down when dynamic objects move non-rigidly relative to the ego vehicle in real-world scenes. We propose Efficient Deformable Temporal Alignment (E-DTA), a lightweight framework that aligns features across time via adaptive deformable sampling. In E-DTA, the sampling radius is adjusted based on feature correlation, expanding the search range for moving objects while keeping it tight in stable regions to reduce background noise. Furthermore, we introduce an adaptive gating mechanism that fuses multi-frame features by balancing the current observation against historical context on a per-location basis. On the nuScenes benchmark, E-DTA with a ResNet-50 backbone achieves 38.5% mAP and 50.1% NDS, and reduces the velocity error (mAVE) to 0.332. These results indicate that E-DTA improves temporal alignment while remaining efficient for autonomous driving perception.
Institution: Shanghai Polytechnic University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7201898566