ASCNet: 3D object detection from point cloud based on adaptive spatial context features

Abstract

This paper presents a novel two-stage 3D point cloud object detector named ASCNet for autonomous driving. Most current works project 3D point clouds to 2D space, whereas the quantization loss in the transformation is inevitable. A Pillar-wise Spatial Context Feature Encoding (PSCFE) module is proposed in the paper to drive the learning of discriminative features and reduce the detailed information loss. The inhomogeneity that existed in 3D object detection from the point clouds, such as the inconsistent number of points in the pillars, the diverse size of Regions of Interest (RoI), should be treated wisely due to the sparsity and the individual specificity. We introduce a length-adaptive RNN-based module to solve the inhomogeneity. A novel backbone combining encoder-decoder and shortcut connection is designed in the paper to learn the multi-scale features for 3D object detection. Additionally, we utilize multiple RoI heads and class-wise NMS to deal with the class imbalance in scenes. Extensive experiments on the KITTI dataset demonstrate that our algorithm achieves competitive performance in 3D bounding box detection and BEV detection.