Based on papers: "OmniMVS: End-to-End Learning for Omnidirectional Stereo Matching" and "End-to-End Learning of Geometry and Context for Deep Stereo Regression"
Modified for binocular system like Intel REALSENSE T265
Codes were heavily borrowed from matsuren/omnimvs_pytorch
Dataset is generated by using Habita-Sim + Replica-Dataset
Because the realization of the double-sphere fisheye camera model in Habitat-Sim is a cubemap hacking (check out here), the results are therefore not accurate. As we can see the cubemap's trace is remarkable:
