AttentiveNet: Detecting Small Objects for LiDAR Point Clouds by Attending to Important Points

With the development of autonomous driving, object detection in LiDAR point clouds is receiving increasing attention. Though many excellent works have been proposed, small object detection still remains challenging because insufficient points of these small objects are scanned, leaving limited information for detectors. In this paper, we propose AttentiveNet, aiming to facilitate small object detection by enriching point information of small objects and focusing on the most important points. The proposed Small object Enhancement Module (SEM) effectively alleviates the issues of insufficient information provided by scanned small objects through an "enriching-attending" strategy. The "enriching-attending" strategy operates by first supplementing points for small objects and then extracting features using a 3D spatial attention mechanism, which helps avoid noise from some inappropriately added points. Furthermore, we propose the Local Attentive 2D Backbone (LAB) to perform self-attention on projected 2D feature maps, catering to the locality of small objects. Extensive experimental results on the KITTI dataset substantiate that the AttentiveNet outperforms existing popular 3D object detection algorithms in detecting 3D small objects in LiDAR point clouds.