jiang-cx/h-omnistereo

H-OmniStereo is a computer vision research project from a group at HKUST, written up in a paper on arXiv as 2605.14963. It is about stereo matching on 360-degree omnidirectional cameras, which is the problem of figuring out how far away each part of a scene is when you have two panoramic cameras instead of two regular ones pointing forward. The motivation, in plain terms: regular stereo matching has gotten very good because there are large training datasets and pretrained networks that can estimate properties like surface orientation from a single image. When you try to copy that pipeline over to 360-degree top-and-bottom camera rigs, two things go wrong. There are not enough public 360 stereo datasets to train on, and the pretrained helpers from regular cameras get confused by the heavy warping that happens when you flatten a sphere into a rectangular image. The paper proposes two fixes. First, the authors built a synthetic dataset with over 2.8 million pairs of top-and-bottom 360 stereo images, which is large enough to train a competitive model. Second, they trained a new normal-estimation network, a model that predicts which way each pixel of the scene is facing, operating directly in the 360 image space with a heading-aligned coordinate system so that the spherical distortion does not throw it off. That gives the stereo matcher a stable geometric hint to work with. The headline claim is that the resulting model is zero-shot: it does well on datasets it was not trained on, and it generalises to real-world setups from off-the-shelf consumer 360 cameras using a single model. The authors say both the model and the dataset will be open-sourced. The current state of the repo is more of a project landing page than a code release. There is a logo, a link to the arXiv paper, a link to a YouTube demo video, the abstract, an author contact email, and the BibTeX citation. There is no installation guide, no Python code, and no dataset download link visible in the README yet.