Free Performance Boost for Voting-based 3D Object Detection in Point Clouds
Rong Yang, Mengjie Qin, Zheyuan Lin, Chunlong Zhang, Te Li, Minhong Wan · 2023
Recently, detecting objects from 3D point clouds based on the voting mechanism has received increasing attention. Various improvements such as RGB feature embedding, context information fusion and bounding box refinement have been developed to further boost the performance of voting-based models since VoteNet was proposed. In this work, we revisit the VoteNet model and point out the significant negative impact brought by votes from background, which has never been discussed properly in previous research as far as we know. In order to suppress votes from the background to obtain purer vote clusters, we redesign the loss function for the training of voting module of voting-based 3D detector. Experiments on the challenging SUN RGB-D dataset show that the new loss function can significantly boost the detection performance of existing state-of-the-art voting-based 3D detectors with no additional computational overhead.