The current fusion-based methods transform LiDAR data into bird’s eye view (BEV) representations or 3D voxel, leading to information loss and heavy computation cost of 3D convolution. In contrast, we directly consume raw point clouds and perform fusion between two modalities. We employ the concept of region proposal network to generate proposals from two streams, respectively. In order to make two sensors compensate the weakness of each other, we utilize the calibration parameters to project proposals from one stream onto the other. With the proposed multi-scale feature aggregation module, we are able to combine the extracted regionof-interest-level (RoI-level) features of RGB stream from different receptive fields, resulting in fertilizing feature richness. Experiments on KITTI dataset show that our proposed network outperforms other fusion-based methods with meaningful improvements as compared to 3D object detection methods under challenging setting.
Optimization of the PointPillars network for 3D object detection in point clouds
