DATNet-RS: Domain-Adaptive Temporal Attention with Residual Shrinkage and Online PSO for Robust Object Detection in Adverse Weather

Engineering notes

Key topics: autonomous driving, object detection, real-world driving, control. 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

Extreme weather conditions are a serious setback to the reliability of object detection in real-world driving conditions. This paper introduces DATNet-RS, a domain adaptive detection system that runs effectively in these conditions and does not require retraining or labelled weather data. A six-component multi-scale attention module is proposed, including local and global versions of channel, spatial, and temporal attention, which, in combination, reduce the noise-enhancing feature channels, highlight spatially consistent object regions, and capitalize on the frame-to-frame temporal continuity. Second, there is a common residual shrinkage denoising block on all levels of the feature pyramid to reduce the low-amplitude noise activations caused by weather, but not the structurally informative responses. Third, a gradient-free online inference-time adaptation scheme is added, where a small nine-dimensional parameter vector - controlling all attention magnitudes and shrinkage thresholds - is jointly optimized by the use of Particle Swarm Optimization (PSO). Experiments demonstrate that DATNet-RS continuously improves the performance compared to the baseline in all four adverse conditions and increases average mAP@0.5 to 75.1% (+4.3 percentage points) and average mAP@0.5:0.95 to 44.2% (+2.9 percentage points), in addition to maintaining real-time performance of GPU inference at about 50 FPS. On mAP at 0.5, the improvement of per-condition is between +3.9% (night) to +4.9% (rain). The assessment is done by isolating the contribution of each component in a seven-configuration ablation study and making comparisons between these and YOLOv9, RT-DETR, and Deformable DETR position DATNet-RS, a highly competitive real-time detector in the evaluated set.

Links and sources

• PDF from original source

Comments