A vehicle detection method based on disparity segmentation

The detection of small objects has always been one of the key challenges in vehicle detection. In this work, a standard for dividing the object more accurately than traditional methods is presented. Based on the division standard of disparity segmentation, we propose a novel multi-scale detection network aiming to reduce the transmission of redundant information between each scale. We divide the objects by depth, which is the distance from the object to the viewpoint. Then, a multi-branch architecture providing specialized detection for objects of each scale separately is constructed. Through ablation experiments, our method achieves an increase of 1.63 to 2.01 mAP compared with the baseline method. On the KITTI dataset, our method combined with state-of-arts achieves an increase of 3.54 mAP for small objects and 0.79 mAP for medium objects.