Overview

The dataset consists 5165 image pairs and corresponding disparity maps, where 4156 image pairs are used for training, and 1009 image pairs are used for testing. The images are extracted from Apollo dataset. Ground truth has been acquired by accumulating 3D point clouds from Lidar and fitting 3D CAD models to individually moving cars (obtained from 3d car instance understanding dataset). The dataset contains varying traffic conditions with heavy occlusion, which are very challenging.