Real-time Hierarchical Soft Attention-based 3D Object Detection in Point Clouds

Deep neural network-based 3D object detection in LiDAR point clouds has achieved excellent performance in various applications including autonomous driving and robot vision. However, achieving high accuracy in real-time is paramount in time-critical applications. We propose a real-time Hierarchical Soft Attention Network (HSAN) to employ soft attention in the backbone of the original network to increase the detection accuracy without slowing down its inference speed. The proposed HSAN applies a hierarchical structure on the baseline network to combine features at different scales to obtain rich and fine-grained information and utilizes the characteristic of a layered attention structure to give more attention to the correct regions of target objects. Our proposed system improves the baseline network and achieves comparable detection results in terms of detection accuracy and inference speed when compared with peer state-of-the-art systems on the KITTI validation 3D detection benchmark.