Residual MBConv Submanifold Module for 3D LiDAR-based Object Detection

Abstract

In LiDAR-based point cloud, objects are always represented as 3D bounding boxes with direction. LiDAR-based object detection task is similar to image-based task but comes with additional challenges. In LiDAR-based detection for autonomous vehicles, the size of 3D object is significant smaller compared with size of input scene represented by point cloud, thus conventional 3D backbones cannot effectively preserve detail geometric information of object with only a few points. To resolve this problem, this paper presents a MBConv Submanifold module, which is simple and effective for voxel-based detector from point cloud. The novel convolution architecture introduces inverted bottleneck and residual connection into 3D sparse backbone, which enable detector to learn high dimension feature from point cloud. Experiments shows that MBConv Submanifold module bring consistent improvement over the baseline method: MBConv Submanifold achieves the AP of 68.03% and 54.74% in the moderate cyclist and pedestrian category on the KITTI validation benchmark, surpass the baseline method significantly. Our code and pretrained models are available at: https://github.com/s1mpleee/ResMBSubmanifold.