Spatially-guided Chunk-wise Reweighting Transformer for 3D Object Detection in Autonomous Vehicles

In computer vision, accurately detecting objects in three-dimensional (3D) scenes is indispensable for intelligent transportation, robotics, vision applications, and augmented reality. While self-attention mechanisms have demonstrated remarkable success in enhancing feature representations and capturing long-range dependencies, their application to 3D object detection remains challenging due to inherent limitations to capturing intricate contextual dependencies among points to refining 3D proposals. This paper proposes a high-quality Chunk-wise Reweighting Transformer for 3D Object Detection (CRT3D) architecture explicitly tailored for 3D object detection tasks. CRT3D comprises three modules: self-attention-based encoder-decoder for effective modelling of spatial context; channel-based Reweighting captures temporal dependencies; and Chunk-wise Reweighting improves computational efficiency to efficiently handle sparse point cloud data, advancing real-time 3D object detection in autonomous vehicles. Our CRT3D model can effectively balance the trade-off between capturing local and global context, leading to more informative and contextually rich representations. Through extensive experiments on the benchmark dataset KITTI [1], we demonstrate the efficacy of CRT3D in achieving better performance by 3.49% and 4.25% in 3D AP with PointPillar and PointRCNN backbone, respectively. CRT3D enhances 3D object detection framework accuracy, particularly at medium range and occlusion cases, while minimally impacting real-time performance and model parameters.