README.md

metadata

license: mit

HiSpatial Training Data

We open-source a large subset of the training data used in HiSpatial, consisting of 1.2M images with rich spatial annotations:

Split	Images	Source	Description
In-the-Wild-o365	~1M	Objects365	Internet images with MoGe2-estimated metric monocular point clouds and multi-level spatial QA annotations (orientation, spatial relation, distance, depth comparison, object size, perspective taking, etc.)
3D-Labeled	~200K	CA-1M	3D scene images with ground-truth camera intrinsics, metric monocular point clouds, and QA annotations:(orientation, perspective_taking, spatial_relation, distance, problem_solving etc.)

📄 Paper: HiSpatial

For data loading and usage, please refer to the dataloader in https://github.com/microsoft/HiSpatial.