---
pretty_name: SpatioLM Depth Benchmark
license: other
task_categories:
- depth-estimation
- visual-question-answering
tags:
- spatial-reasoning
- metric-depth
- multi-view
- camera-geometry
- spatiolm
size_categories:
- 1K<n<10K
---

# SpatioLM Depth Benchmark

This dataset contains the processed depth and spatial-relation benchmark used by
[SpatioLM](https://github.com/xiaomi-research/spatio-lm). It is distributed in
the Hugging Face `DatasetDict.save_to_disk` format to remain directly compatible
with the evaluation tasks included in the SpatioLM repository.

## Dataset structure

| Split | Examples | Description |
| --- | ---: | --- |
| `single_view` | 1,600 | Single-image metric depth samples derived from SUN RGB-D, NYUv2, and Waymo. |
| `multi_view` | 1,492 | Paired-view metric depth samples derived from NRGBD, ScanNet v2, and KITTI. |
| `relate_task` | 2,800 | Depth-related reasoning over speed/time, two-point distance, camera pose, and cross-view geometry. |

Images and depth maps are embedded in the Arrow files. Depending on the split,
examples additionally contain sampled pixel coordinates, camera-space depth,
Euclidean distance, camera intrinsics, and camera poses.

## Download and load

Download the complete repository without changing its directory structure:

```bash
hf download edatai/spatiolm-depth \
  --repo-type dataset \
  --local-dir data/eval/spatiolm_depth
```

Load it with `datasets.load_from_disk`:

```python
from datasets import load_from_disk

dataset = load_from_disk("data/eval/spatiolm_depth")
print(dataset)
```

The repository is intentionally not converted to Hub-native Parquet. Use
`load_from_disk`, not `load_dataset`, to preserve compatibility with the
SpatioLM evaluation task definitions.

## SpatioLM evaluation tasks

| Task | Split |
| --- | --- |
| `spatiolm_depth_sv` | `single_view` |
| `spatiolm_depth_mv` | `multi_view` |
| `spatiolm_depth_mt` | `relate_task` |

See the [SpatioLM repository](https://github.com/xiaomi-research/spatio-lm) for
the model adapter, metrics, and evaluation commands.

## Source data and licensing

This is a processed benchmark derived from SUN RGB-D, NYUv2, Waymo, NRGBD,
ScanNet v2, and KITTI. The images, depth maps, calibration data, and derived
annotations remain subject to the licenses and terms of their respective source
datasets. No single permissive license is asserted for the combined benchmark.
Users are responsible for complying with all applicable upstream terms.

## Citation

```bibtex
@inproceedings{wu2026spatiolm,
  title={SpatioLM: Towards General Physical Spatial Intelligence in Vision-Language Models},
  author={Wu, Jing and Wu, Jianhua and Guan, Jiayi and Chen, Jiahong and Lu, Jinghui and Ye, Hangjun and Gao, Bingzhao and Chen, Long},
  booktitle={International Conference on Machine Learning (ICML)},
  year={2026},
  note={To appear},
  eprint={2608.01899},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2608.01899}
}
```