MoCha-Stereo: Motif Channel Attention Network for Stereo Matching

† Figure 1. Motivation. Addressing the issue of geometric structure loss in feature channels arising from deep learning. (From left to right: Input image, visualization of a Normal Channel, visualization of a Motif [40] Channel.) Due to the fuzziness of geometric edges in certain channels, achieving accurate matching of stereo image edges is a challenging problem. MoCha-Stereo guides ordinary channels to focus on edge features through motif channels, achieving more accurate detail matching. Motif Channel refers to channel that composed of repeatedly occurring geometric contours. The regions delineated by the yellow border represent the magnified details.