The IndoReal-MVS dataset is currently in the stages of preparation and will soon be released.
IndoReal-MVS offers an advanced tool for multi-view stereo (MVS) research. Developed with cutting-edge computer graphics, it mirrors real-world phenomena such as dynamic lighting, reflections, and shadows across 31 diverse indoor environments. This dataset features 21,504 (480x640) images, encompassing living areas, workspaces, and staircases, complete with detailed ground truth depth, normal maps, and camera parameters for both supervised and unsupervised learning paradigms.
Split into training (24 scenes, 14,473 images) and testing (7 scenes, 7,031 images) subsets, it provides comprehensive coverage for robust model evaluation in a variety of indoor settings. This upcoming release hopes to enhance realism and applicability for depth estimation and scene understanding projects.