Practical Deep Stereo (PDS): Toward applications-friendly deep stereo\n matching

Abstract

End-to-end deep-learning networks recently demonstrated extremely good\nperfor- mance for stereo matching. However, existing networks are difficult to\nuse for practical applications since (1) they are memory-hungry and unable to\nprocess even modest-size images, (2) they have to be trained for a given\ndisparity range. The Practical Deep Stereo (PDS) network that we propose\naddresses both issues: First, its architecture relies on novel bottleneck\nmodules that drastically reduce the memory footprint in inference, and\nadditional design choices allow to handle greater image size during training.\nThis results in a model that leverages large image context to resolve matching\nambiguities. Second, a novel sub-pixel cross- entropy loss combined with a MAP\nestimator make this network less sensitive to ambiguous matches, and applicable\nto any disparity range without re-training. We compare PDS to state-of-the-art\nmethods published over the recent months, and demonstrate its superior\nperformance on FlyingThings3D and KITTI sets.\n