README.md

---
license: cc-by-nc-sa-4.0
tags:
  - 3d
  - scene-understanding
  - embeddings
pretty_name: InternScenes SSP records
extra_gated_prompt: >-
  Access is granted on request. This data is derived from InternScenes
  (CC BY-NC-SA 4.0) and inherits its NonCommercial and ShareAlike terms.
extra_gated_fields:
  Name: text
  Affiliation: text
  Intended use: text
---

# InternScenes SSP records

Per-object and per-scene encoder records for the [InternScenes](https://huggingface.co/datasets/InternRobotics/InternScenes)
indoor scene corpus: sampled surface geometry paired with DINOv3 patch features.

405,139 records, 425 GB, delivered as ~110 uncompressed tar shards of about 4 GB each.

| group | records | size |
|---|---|---|
| `objects` | 381,365 | 253 GB |
| `scenes/gen` | 14,089 | 133 GB |
| `scenes/real` | 9,583 | 38 GB |
| `scenes/synthetic` | 92 | 0.3 GB |

## Layout

```
objects/objects-000NN.tar          # unpacks to objects/<uid>.npz
scenes-gen/scenes-gen-000NN.tar    # unpacks to scenes/gen/<room>/<id>.npz
scenes-real/...                    # scenes/real/{scannet,matterport3d,3rscan,arkitscenes}/...
scenes-synthetic/...               # scenes/synthetic/<id>.npz
index.csv.gz                       # record -> shard -> bytes, for every record
proj_W_1024x3201.npy               # DINO -> SSP projection
```

`index.csv.gz` lets you locate a single record without unpacking anything:

```python
import pandas as pd
idx = pd.read_csv("index.csv.gz")
idx[idx.record == "scenes/synthetic/0059.npz"]        # -> which shard to download
```

## Record schema

Each `.npz` stores raw components. Scene records concatenate their objects; object records hold one.

| key | shape | dtype | meaning |
|---|---|---|---|
| `mesh_pos` | (N, 3) | float32 | patch centroids, normalized frame |
| `mesh_nrm` | (N, 3) | float32 | patch normals |
| `pooled_cov` | (M, 1024) | float16 | DINOv3 patch features, covered patches only |
| `cov` | (N,) | bool | which patches carry appearance features |
| `owner` | (N,) | int32 | index into the `group_*` arrays (scene records) |
| `group_key` / `group_uid` | (G,) | str | per-object identifiers |
| `group_cat` | (G,) | str | category label (`structure` for walls/floor/ceiling) |
| `group_bbox` | (G, 9) | float32 | 9-DoF box: centre, extent, rotation |
| `scene_C` / `scene_S` | (3,) / () | float32 | normalization offset and scale |
| `n_per_object` | () | int32 | patches per object |
| `ssp_dim`, `ls_pos`, `ls_nrm` | () | int32/float32 | encoder parameters |
| `scene_set`, `scene_name` | () | str | provenance |

`pooled_cov` is compacted to covered patches, so scatter it back before use:

```python
import numpy as np
z = np.load("scenes/synthetic/0059.npz")
pooled = np.zeros((len(z["mesh_pos"]), 1024), np.float32)
pooled[z["cov"]] = z["pooled_cov"]
```

World coordinates: `mesh_pos * scene_S + scene_C`.

`proj_W_1024x3201.npy` is a seeded draw and reproducible independently:

```python
W = (np.random.default_rng(0).standard_normal((1024, 3201)) / np.sqrt(1024)).astype(np.float32)
```

## Known gaps

Five synthetic scenes have no record (`0004`, `0026`, `0088`, `0090`, `0097`).

## Provenance and license

Source scenes: [InternScenes](https://huggingface.co/datasets/InternRobotics/InternScenes),
CC BY-NC-SA 4.0. This derived data is released under the same license, and NonCommercial and
ShareAlike terms carry over. Cite InternScenes if you use it:

```bibtex
@inproceedings{InternScenes,
  title={InternScenes: A Large-scale Interactive Indoor Scene Dataset with Realistic Layouts},
  author={Lin, Weipeng and Cao, Peizhou and Jin, Yichen and Li, Luo and Cai, Wenzhe and
          Lin, Jingli and Lyu, Zhaoyang and Wang, Tai and Dai, Bo and Xu, Xudong and Pang, Jiangmiao},
  booktitle={arXiv},
  year={2025}
}
```

Appearance features are patch embeddings from
[DINOv3 ViT-L/16](https://huggingface.co/facebook/dinov3-vitl16-pretrain-lvd1689m)
(`facebook/dinov3-vitl16-pretrain-lvd1689m`) and are subject to the DINOv3 license.