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.
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.
