3D Object Detection Based on Multi-view Adaptive Fusion

Abstract

Aiming at the problem that multi-view features are difficult to fuse effectively, a multi-view feature adaptive fusion 3D object detection framework is proposed, and new solutions are proposed in two aspects: depth feature fusion and loss function design. It mainly cooperates the bird's-eye view and cylindrical view, carries out adaptive feature fusion on the premise of considering the interaction between views and the contribution of different view features to the detection task, and improves the importance of network learning structure information and local features through the information of two additional tasks: foreground classification and central regression, At the same time, the loss calculation is optimized in the detection process to improve the regression effect of the target boundary box. Experiments on KITTI dataset show that this method achieves higher performance in all single-stage fusion methods, is better than most two-stage fusion methods, and achieves a good balance between speed and accuracy on KITTI benchmark.