Aggregate and Excitate Sparse Spatial Feature: Single-Stage 3D Object Detector from Point Clouds

In order to solve the problems of fixed receptive field and single feature scale in the single-stage voxel-based 3D object detectors, which cause insufficient learning of point clouds features and bottleneck in detection performance, a voxel-based single-stage 3D object detection model is proposed which can be trained end-to-end. First, the multi-scale sparse spatial feature aggregation module is used to aggregate the features of point clouds at different sparse spatial scales, so that the features can fully preserve the spatial information of point clouds. Then, the detector performs hierarchical learning of features through multi-scale receptive fields by feature hierarchical excitation module, which can strengthen the representation of the features, and reduce the influence of noise information to make the detector more robust. Finally, the features are fed into the detection head for classification and regression of candidate boxes. Extensive experiments on the public autonomous driving dataset KITTI compared with other mainstream single-stage 3D object detectors, including 3 categories of 9 difficulty levels, demonstrate that, the average accuracy of the proposed model is significantly improved in 5 levels, especially for objects with sparse point clouds. The experimental results show that the proposed model can fully extract the spatial information and effectively learn the multi-scale features of the point clouds.