HITNet: Hierarchical Iterative Tile Refinement Network for Real-time Stereo Matching

This paper presents HITNet, a novel neural network architecture for real-time\nstereo matching. Contrary to many recent neural network approaches that operate\non a full cost volume and rely on 3D convolutions, our approach does not\nexplicitly build a volume and instead relies on a fast multi-resolution\ninitialization step, differentiable 2D geometric propagation and warping\nmechanisms to infer disparity hypotheses. To achieve a high level of accuracy,\nour network not only geometrically reasons about disparities but also infers\nslanted plane hypotheses allowing to more accurately perform geometric warping\nand upsampling operations. Our architecture is inherently multi-resolution\nallowing the propagation of information across different levels. Multiple\nexperiments prove the effectiveness of the proposed approach at a fraction of\nthe computation required by state-of-the-art methods. At the time of writing,\nHITNet ranks 1st-3rd on all the metrics published on the ETH3D website for two\nview stereo, ranks 1st on most of the metrics among all the end-to-end learning\napproaches on Middlebury-v3, ranks 1st on the popular KITTI 2012 and 2015\nbenchmarks among the published methods faster than 100ms.\n