Matching cost function using robust soft rank transformations

Stereo correspondence is a challenging task because stereo images are affected by many factors such as radiometric distortion, sun and rain flares, flying snow, occlusions and object boundaries. However, most of the existing stereo correspondence methods use simple matching cost functions. As a result, their performance is degraded significantly when operating with real-world stereo images whose intensities of corresponding pixels can be arbitrarily transformed. In this study, the authors propose a novel matching cost function based on the order relations between pixel pairs that can operate accurately under various conditions of transformed intensities between stereo images. The proposed matching cost function is an improvement of the soft rank transform (SRT) and can tolerate local, monotonically non-linear changes in intensities between the left and right images. The proposed function significantly reduces the error rate from 24.7 to 12.7% in the Middlebury dataset, and from 19.8 to 7.1% in the KITTI dataset as compared with the SRT. The qualitative and quantitative experimental results obtained using stereo images in different datasets under various conditions show that the proposed matching cost function outperforms the state-ofthe- art matching cost functions in indoor and outdoor stereo images.