File access utilities for the data published with the IROS 2021 submission ORStereo: Occlusion-Aware Recurrent Stereo Matching for 4K-Resolution Images.
We collect 100 4K-resolution stereo images with true disparities and occlusion masks. This dastaset is used for evaluating our model's performance on 4K-resolution images. It is collected by AirSim in the Unreal engine. Get the dataset here.
All utility functions are in file_access_utils.py. The utility depends on OpenCV. Use pip install opencv-python to install OpenCV. The utility has been tested with Python3.
The disparities are saved as 4-channel PNG files. We provide utility functions to read a single disparity file and convert the data into single-precision floating-point values (numpy.float32).
The occlusions are saved as single-channel PNG files. There is a function for reading an occlusion file as a boolean image.
For more details about the paper, please refer to the project webpage.
Created by Yaoyu Hu <yaoyuh@andrew.cmu.edu>