ACF4D: Alignment and Consistency Guided Temporal Fusion for Multi-View 3D Object Detection

Engineering notes

Experimental results on the nuScenes benchmark demonstrate that our proposed method achieves 51.5% mAP and 60.2% NDS.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Abstract Effective temporal feature fusion is essential for improving detection accuracy in image-based Bird’s-Eye View (BEV) perception tasks. However, achieving robust temporal feature fusion faces significant challenges due to feature misalignment caused by calibration errors and localization drift, and inaccuracies in depth estimation, all of which can lead to degraded detection performance. To overcome these challenges, we propose ACF4D, a novel temporal fusion framework designed for multi-view 3D object detection. We introduce a Coarse-to-Fine Temporal Feature Alignment Module (CF-Align) aimed at mitigating temporal misalignment. Additionally, to improve the accuracy of depth estimation, we propose a Temporal Feature Consistency Loss (TFCL), which constrains the discrepancies between features across aligned frames. Furthermore, we construct a Temporal Fusion Network (TFN) utilizing depthwise separable convolution to achieve efficient multi-frame feature fusion while maintaining real-time performance. Experimental results on the nuScenes benchmark demonstrate that our proposed method achieves 51.5% mAP and 60.2% NDS. Moreover, the model operates at 20.4 FPS, demonstrating an effective balance between detection accuracy and computational efficiency.

Links and sources

• PDF from original source

Comments