---
license: other
license_name: hku-academic-non-commercial
license_link: https://huggingface.co/datasets/rayruiyang/GeoUG_data#publisher-permitted-use-and-contact
language:
- en
task_categories:
- visual-question-answering
size_categories:
- n<1K
pretty_name: GeoUG Spatial RL Data
tags:
- geoug
- spatial-reasoning
- reinforcement-learning
- metric-depth
configs:
- config_name: default
  data_files:
  - split: train
    path: train_geoug_rl.parquet
---

# GeoUG Spatial RL Data

**Spatial-policy training data with embedded images, camera parameters, and generated-depth caches.**

This repository hosts the 800-sample training release for GeoUG spatial-policy
reinforcement learning. All required per-sample training assets are stored in
one parquet: `train_geoug_rl.parquet` (2,019,132,360 bytes; 1.88 GiB).

[Project page](https://yangr116.github.io/geoug-project-page/) ·
[GeoUG-ColdStart](https://huggingface.co/rayruiyang/GeoUG-ColdStart) ·
[GeoUG-Spatial](https://huggingface.co/rayruiyang/GeoUG-Spatial) ·
[Download parquet](https://huggingface.co/datasets/rayruiyang/GeoUG_data/resolve/main/train_geoug_rl.parquet)

## Sources and construction

The release combines spatial questions derived from the training portions of
[ScanNet++](https://scannetpp.mlsg.cit.tum.de/scannetpp/) and
[nuScenes](https://www.nuscenes.org/nuscenes). Source identifiers are retained
in each row for provenance. Original images and source scene annotations belong
to their respective dataset creators. GeoUG adds the spatial-policy training
examples and packages their camera information and generated-depth caches.

| `source` | Samples |
|---|---:|
| `scannetpp_train` | 400 |
| `nuscenes_train` | 400 |

| Preserved task category | Samples |
|---|---:|
| `multi_object_closest_among` | 200 |
| `camera_closest_among` | 200 |
| `multi_object_closer_to` | 200 |
| `location_closer_to_camera` | 200 |

Category names are grouping identifiers; use the actual question text to
interpret closer/farther wording. Answers may be A, B, C, or D depending on
the options. The generated-depth arrays are model-generated training caches,
not ground-truth depth labels from the source datasets.

## Download and inspect a sample

```python
import io
import json
import numpy as np
import pyarrow.parquet as pq
from PIL import Image
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="rayruiyang/GeoUG_data",
    repo_type="dataset",
    filename="train_geoug_rl.parquet",
)
row = next(pq.ParquetFile(path).iter_batches(batch_size=1)).to_pylist()[0]
image = Image.open(io.BytesIO(row["image"])).convert("RGB")
depth = np.load(io.BytesIO(row["depth_npy"]), allow_pickle=False)
camera = row["camera"]
options = json.loads(row["options"])
print(row["question"], row["answer"], depth.shape)
```

## Schema

| Field | Meaning |
|---|---|
| `id`, `image_key` | Question and image identifiers |
| `image` | Encoded image bytes |
| `question`, `answer`, `options` | Question, target answer, and JSON-encoded option mapping |
| `category`, `source` | Task grouping and source provenance |
| `illusion`, `illusion_strength` | Construction-time shortcut/illusion diagnostics |
| `camera.intrinsics` | 3 × 3 camera intrinsic matrix |
| `camera.image_wh` | Width and height associated with the camera calibration |
| `depth_npy` | Original NumPy depth-cache file bytes, embedded without precision conversion |

Camera calibration dimensions may differ from the encoded image or depth-map
resolution. Use the calibration dimensions when rescaling intrinsics, as done
by the GeoUG geometry tools. Do not assume that all three resolutions match.

## Use for RL training

From the GeoUG code checkout, restore the existing training layout:

```bash
python scripts/pack_data_parquet.py unpack-rl \
  --input /absolute/path/to/train_geoug_rl.parquet \
  --output /absolute/path/to/rl_train
```

This creates:

```text
rl_train/
├── train.parquet
├── camera_table.json
└── depth-cache/
```

Set `RL_DATASET=/absolute/path/to/rl_train` and
`DEPTH_CACHE_DIR=/absolute/path/to/rl_train/depth-cache`, then follow
`docs/RL_TRAINING.md`. Use GeoUG-ColdStart as the initialization checkpoint.
No separate camera-table or depth-cache download is required.

This release contains **training samples only**. The internal RL evaluation
subset, geometric co-training / VLM manifests, and cold start training data
are not included. Prepare validation/evaluation inputs separately.

## File integrity

```text
SHA256: 425c12e2eff223135cc3c159e70ebd2318b4e888f58dc7f4dfd243fc67673476
```

## Publisher, permitted use, and contact

The following notice applies to this GeoUG data release:

It is published by The University of Hong Kong (HKU) and is licensed for non-commercial use only. Its use is strictly limited to academic research and educational purposes. Please contact [**rayyang0116@connect.hku.hk**](mailto:rayyang0116@connect.hku.hk) if you have any concerns regarding the data.

The original datasets are credited in the sources section. This release does
not transfer ownership of upstream images or annotations; their original
licenses, attribution requirements, and terms of use remain applicable.