Evaluating and Improving Disparity Maps Without Ground Truth

Disparity maps are integral to the fields of computer vision and computer graphics, with many applications in robotics, path planning, computerized drawing, 3D viewing, and more. Stereo algorithms can be used to synthesize disparity maps from stereo image pairs. Unfortunately, most methods to assess disparity map quality rely on a ground-truth disparity map, which is not always available. This paper introduces a method to evaluate disparity map generation algorithms that does not require ground truth, and is instead based on the quality of the disparity maps themselves. This novel scoring metric assesses each pixel in a given disparity map and then provides an overall score for the disparity map. The score is both a visual representation—where each color encodes an error or potential error—and a numerical score. This novel scoring metric is used to compare two top-ranking disparity map generators in different scenarios. Our scoring method yields results consistent with intuition and also uncovers, and helps in resolving, an issue with one generator. Finally, RANSAC plane-fitting is used to automatically inpaint unknown regions and thereby improve the usability of disparity maps in applications such as non-photorealistic rendering, layer-based stereo methods, and warping.