| --- |
| pretty_name: Pointcept2 processed point cloud datasets |
| license: other |
| license_name: per-source-dataset-terms |
| tags: |
| - point-cloud |
| - 3d |
| - scene-understanding |
| extra_gated_prompt: >- |
| This repository redistributes processed versions of third-party 3D datasets. Each dataset keeps |
| its original terms of use (linked below). By requesting access you confirm that you have accepted |
| the terms of every dataset you download from here and that you will use the data for |
| non-commercial research only. |
| extra_gated_fields: |
| Name: text |
| Affiliation: text |
| I have accepted the terms of use of the source datasets I will download: checkbox |
| --- |
| |
| # Pointcept2 processed point cloud datasets |
|
|
| Indoor scene datasets processed for [Pointcept2](https://github.com/Pointcept): one `.safetensors` |
| file per scene, a `dataset.json` card per dataset (fields, dtypes, roles, units, classes, |
| provenance) and an `index.json` manifest (splits, point counts, byte sizes, sha256 of every file). |
| Pointcept2 reads them directly; any safetensors reader can too. |
|
|
| | folder | source | scenes | content | notes | |
| |---|---|---|---|---| |
| | `scannet/` | ScanNet v2 | 1,613 (train 1,201 / val 312 / test 100) | coord (m, f32), color (u8), normal (f32), segment20, segment200 (i16, -1 ignore), instance (i32) | Pointcept's preprocessing ported bit-identically (same labels and order) | |
| | `structured3d-full/` | Structured3D | 21,821 rooms (all splits) | coord, color, normal, segment, instance | derived version: Pointcept's rooms pre-sampled to one point per 1.5 cm cell | |
| | `s3dis/` | S3DIS (Stanford 2D-3D-S) | 272 rooms (Area_1 to Area_6) | coord, color, normal, segment (13 classes), instance | Pointcept's processed rooms (aligned, with normals), stored as-is | |
| | `arkitscenes/` | ARKitScenes (3DOD meshes) | 5,047 scenes (Training / Validation) | coord, color, normal | Pointcept's processed scenes, stored as-is; no labels | |
| | `hm3d/` | HM3D v0.2 | 11,493 rooms (train / val) | coord, color, normal; segment, instance where HM3D-Sem has them | Pointcept's processed rooms, stored as-is | |
| | `scannetpp/` | ScanNet++ v2 | 956 scenes (train 856 / val 50 / test 50) | coord, color, normal, segment, instance | Pointcept's processed meshes, full resolution | |
| | `scannetpp-g0.01/` | ScanNet++ v2 | 956 scenes | as `scannetpp/` | derived version: one point per 1 cm cell, for training | |
|
|
| A folder is complete when its `index.json` is present. Each folder's `dataset.json` records the exact |
| source, its hashes and the processing. |
|
|
| **Tar shards.** The hub allows at most 10,000 files in one directory, so a split with more samples is |
| stored as uncompressed tar shards next to the other splits: `structured3d-full/train-000.tar`, ... and |
| `hm3d/train-000.tar`, ... (about 4 GB each, samples in index order). Unpack them inside the dataset |
| folder to get the layout `index.json` lists, then delete the shards: |
|
|
| ```bash |
| cd structured3d-full && for f in train-*.tar; do tar -xf "$f" && rm "$f"; done |
| ``` |
|
|
| Every sample's sha256 is in `index.json`. |
|
|
| ## Terms of use |
|
|
| The data remain under the terms of their original datasets; this repository grants no additional |
| rights. Read and accept them before use: |
|
|
| - ScanNet: https://kaldir.vc.in.tum.de/scannet/ScanNet_TOS.pdf |
| - Structured3D: https://structured3d-dataset.org/ (terms of use on the download form) |
| - S3DIS / 2D-3D-S: http://buildingparser.stanford.edu/dataset.html |
| - ARKitScenes: https://github.com/apple/ARKitScenes/blob/main/LICENSE |
| - HM3D: https://matterport.com/habitat-matterport-3d-research-dataset (Matterport's terms) |
| - ScanNet++: https://kaldir.vc.in.tum.de/scannetpp/ |
| |
| ## Reading a sample |
| |
| ```python |
| from safetensors.numpy import load_file |
| scene = load_file("scannet/train/scene0000_00.safetensors") |
| coord, color = scene["coord"], scene["color"] # (N, 3) float32 metres, (N, 3) uint8 |
| ``` |
| |
| ## Citation |
| |
| Cite the original datasets (ScanNet, Structured3D, ...) and Pointcept. |
| |
