3D object detection in disparity space using context guided stereo matching

Abstract

Most of the recent 3D object detection algorithms are based on LiDAR point cloud and algorithms based on images still lacks behind in performance. This is mostly attributed to the less accurate depth estimation from images compared to LiDAR. However, LiDAR are expensive sensors and cameras offers a more efficient alternative if we can construct a reliable algorithm. In this study, we investigate a novel 3D object detection algorithm that utilizes stereo images. The algorithm makes detection on the disparity space images constructed from stereo matching. The features of the images in the disparity space are already computed during the stereo matching algorithm. This way, we expand the detection space from 2D image to 2.5D space while sharing computation with stereo matching algorithm to keep computation cost lower. For that, we also built a memory efficient stereo matching algorithm by modifying the cost volume construction step of existing algorithm. Our stereo matching algorithm is shown to be faster and memory efficient compared to previous methods.