NV-DETR: A Lightweight and Accurate End-to-End Object Detector for Autonomous Driving

Engineering notes

Key topics: autonomous driving, end-to-end, object detection, kitti. See the paper for implementation details and experimental results.

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

Accurate and efficient object detection is vital for intelligent transportation and autonomous driving, especially in complex urban environments. We propose Nano and Veracity Detection Transformer (NV-DETR), a lightweight, highperformance detection framework built on RT-DETR. NV-DETR integrates a WT Block for multi-scale frequency enhancement, a parameter-free SPD-Conv for efficient spatial compression and channel expansion, and a Multi-Scale Hierarchical Attention Fusion (MS-HAF) module for adaptive feature integration. On BDD100K and KITTI datasets, NV-DETR reduces parameters by 22.6 % and GFLOPs by 17.7 %, while improving mean average precision (mAP)@0.5:0.95 by $\mathbf{3. 0 \%}$ and $\mathbf{2. 3 \%}$, respectively. For Average Precision (AP)-Small, relative gains reach 6.8 % on BDD100K and 1.8 % on KITTI, demonstrating enhanced object detection capability for small-scale targets in dense, complex traffic scenes, with an optimal balance between accuracy and efficiency.

Links and sources

• Official / arXiv page

Comments