Learning to recover weak signals: A two-stage graph-based framework for all-weather RGB-T object detection

Engineering notes

Extensive experiments on the KAIST and R-LiViT datasets demonstrate that our framework significantly improves robustness across day and night scenarios, reducing the day-night performance gap from 72.4% to 7.4% and substantially enhancing the detection of small objects. Moreover, the method achieves a real-time inference speed of 27.4 FPS, offering a favorable trade-off between accuracy and efficiency for autonomous driving perception in challenging environments.The code is available at: https://github.com/yangjiepry/YOLO-HA-GNN.git .

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

Object detection in complex traffic environments remains challenging due to illumination-induced feature degradation and weak responses from distant small objects. To address these limitations, we propose a two-stage multimodal detection framework that enhances perception under degraded conditions. The first stage employs a Dual-Path Attention Fusion Module (DPAFM) to adaptively integrate RGB and thermal features via learnable gated attention, mitigating modality-specific degradation. The second stage introduces a Hierarchical Attention Graph Neural Network (HA-GNN), which models detection candidates as graph nodes and performs hierarchical relational reasoning over their appearance and spatial relationships to compensate for weakened local cues. Extensive experiments on the KAIST and R-LiViT datasets demonstrate that our framework significantly improves robustness across day and night scenarios, reducing the day-night performance gap from 72.4% to 7.4% and substantially enhancing the detection of small objects. Moreover, the method achieves a real-time inference speed of 27.4 FPS, offering a favorable trade-off between accuracy and efficiency for autonomous driving perception in challenging environments.The code is available at: https://github.com/yangjiepry/YOLO-HA-GNN.git .

Links and sources

Comments