Object-Centric Stereo Matching for 3D Object Detection

Safe autonomous driving requires reliable 3D object detection-determining the6 DoF pose and dimensions of objects of interest. Using stereo cameras to solvethis task is a cost-effective alternative to the widely used LiDAR sensor. Thecurrent state-of-the-art for stereo 3D object detection takes the existingPSMNet stereo matching network, with no modifications, and converts theestimated disparities into a 3D point cloud, and feeds this point cloud into aLiDAR-based 3D object detector. The issue with existing stereo matchingnetworks is that they are designed for disparity estimation, not 3D objectdetection; the shape and accuracy of object point clouds are not the focus.Stereo matching networks commonly suffer from inaccurate depth estimates atobject boundaries, which we define as streaking, because background andforeground points are jointly estimated. Existing networks also penalizedisparity instead of the estimated position of object point clouds in theirloss functions. We propose a novel 2D box association and object-centric stereomatching method that only estimates the disparities of the objects of interestto address these two issues. Our method achieves state-of-the-art results onthe KITTI 3D and BEV benchmarks.