Fig. 20: Flowchart of the deep-learning-based method for extracting depth information and the estimated disparity images using different methods.

a The flowchart of deep-learning-based method for extracting depth information: two network architectures (one tuned for speed, the other for accuracy) are trained to learn the matching cost computation. The output of CNN is applied to initialize the stereomatching cost, followed by a series of postprocessing processes. b, c The input stereo images. d Ground truth. e, g The disparity estimation results using Census³³⁵ and CNN. f, h The disparity errors of (e, g). a–h Adapted from ref. ³³⁴. © 2016 Jure Zbontar and Yann LeCun, Microtome Publishing