README.md

---
pretty_name: AFUN_pool
license: cc-by-nc-sa-4.0
language:
- en
size_categories:
- 100K<n<1M
tags:
- robotics
- affordance
- manipulation
- 3d-motion
- segmentation
---

# AFUN_pool

The full data pool of **AFUN** ([arXiv:2606.02551](https://arxiv.org/abs/2606.02551)):
**183,657 data points** for affordance segmentation and 3D interaction-motion prediction,
i.e. the fitted-curve pool of Table 7 from which the curated
[**AFUN**](https://huggingface.co/datasets/AFUN-dataset/AFUN) training set (44,749) was
selected. AFUN ⊂ AFUN_pool. Same folder format as AFUN, same `trajectory.json` schemas.

It comes in two parts:

| part | data points | what each folder contains |
|---|---:|---|
| **full** | 77,432 | `obs_frame.png`, `obs_frame_depth.npy`, `sam_mask.png`, `trajectory.json` |
| **ego4d** (annotations only) | 106,225 | `sam_mask.png`, `trajectory.json`, `provenance.json` — RGB/depth are rebuilt from your own Ego4D download, see below |

The 106,225 `ego4d` data points are human videos from [Ego4D](https://ego4d-data.org),
whose license allows redistributing annotations but not the video frames themselves.
Everything we produced for them is here; the frame is one script call away.

Of the 223,334 fitted curves in Table 7, 39,677 are not in this release: their affordance
mask was not tracked at the observation frame, or their `trajectory.json` lacks `motion_2d`
(HOI4D).

## Download & extract

```bash
pip install -U huggingface_hub
hf download AFUN-dataset/AFUN_pool --repo-type dataset --local-dir afun_pool
cd afun_pool
for f in data/*.tar.zst; do tar --zstd -xf "$f"; done
```

Download ≈ 363 GiB (of which the `ego4d` part is 0.26 GiB), extracted ≈ 748 GiB. After extraction:

```
afun_pool/
├── manifest.json                          # index — one entry per data point, with `part`
├── reconstruct_ego4d_frames.py            # rebuilds obs_frame.png for the ego4d part
├── <source>/<episode>/<interval>/<cam>/   # 77,432 folders  (part = "full")
│   ├── obs_frame.png                      # RGB frame
│   ├── obs_frame_depth.npy                # float32 H×W depth, millimeters
│   ├── sam_mask.png                       # affordance mask (non-zero = actionable region)
│   └── trajectory.json                    # 3D motion + camera intrinsics
└── ego4d/<episode>/<interval>/<cam>/      # 106,225 folders (part = "ego4d")
    ├── sam_mask.png
    ├── trajectory.json
    └── provenance.json                    # which Ego4D frame this is (+ pixel hash)
```

Loading a data point is identical to AFUN — `manifest.json` entries carry `path`,
`dataset` / `episode_id` / `interval` / `cam`, `language`, `shard`, plus `part`.
`trajectory.json` follows the two AFUN schemas (top-level fields for robot / human sources,
nested `trajectories[]` for `scenefun3d`); see the
[AFUN README](https://huggingface.co/datasets/AFUN-dataset/AFUN#trajectoryjson).

## Rebuilding the Ego4D frames

Each `ego4d` folder's `provenance.json` records the Ego4D video and frame:

```json
{
  "ego4d_video_uid": "0031d268-818c-4ec4-a804-935be610a61a",
  "ego4d_frame_index": 56979,           // 0-based frame in the full_scale video
  "fps": 30.0,
  "image_hw": [1440, 1920],
  "rgb_sha256": "b22fda36…",            // sha256 of the raw H×W×3 uint8 pixels
  "language": "Spread adhesive with the trowel over the underlayment.",
  "vitra_episode_file": "ego4d_other/episodic_annotations/Ego4D_0031d268-…_ep_000859.npy",
  "episode_frame_index": 0,             // index within that VITRA-1M episode
  "depth": { "model": "depth-anything/DA3NESTED-GIANT-LARGE-1.1", "...": "..." }
}
```

1. Get Ego4D access at https://ego4d-data.org and download only the videos you need:

   ```bash
   pip install ego4d av Pillow
   python reconstruct_ego4d_frames.py --pool-root ego4d --list-uids > uids.txt
   ego4d --output_directory ~/ego4d --datasets full_scale --version v2 --video_uid_file uids.txt
   ```

2. Decode the frames into place and verify them against the recorded pixel hashes:

   ```bash
   python reconstruct_ego4d_frames.py --pool-root ego4d --ego4d-root ~/ego4d --verify
   ```

   This writes `obs_frame.png` into every `ego4d/...` folder using the same decoder and
   frame indexing the annotations were made with (PyAV, native resolution, no resizing);
   `--verify` confirms each decoded frame matches `rgb_sha256`.

The intervals were taken from [VITRA-1M](https://github.com/microsoft/VITRA) (MIT), and
`vitra_episode_file` / `episode_frame_index` locate the same frame in its episode files.
Depth maps are not shipped for this part; `provenance.json → depth` gives the settings we
used (Depth Anything 3 streaming video depth, every 4th frame, metric millimeters) if you
want to regenerate them.

## Sources

| key | dataset | part | data points |
|---|---|---|---:|
| ego4d | Ego4D via VITRA-1M (human videos) | ego4d | 106,225 |
| scenefun3d | SceneFun3D | full | 49,706 |
| vitra_epic | EPIC-KITCHENS via VITRA-1M (human videos) | full | 9,155 |
| robomind | RoboMIND | full | 7,077 |
| calvin | CALVIN | full | 3,105 |
| droid | DROID | full | 2,751 |
| robomind2 | RoboMIND 2 | full | 2,356 |
| rh20t_human | RH20T human demos | full | 1,422 |
| rh20t | RH20T | full | 1,044 |
| agibot | AgiBot World | full | 783 |
| rlbench | RLBench | full | 33 |

AFUN's `vitra` source corresponds to `vitra_epic` ∪ `ego4d` here. EPIC-KITCHENS frames are
redistributed under CC BY-NC 4.0 (Damen et al.); Ego4D frames are not redistributed.
The evaluation sets ([**AFUN_eval**](https://huggingface.co/datasets/AFUN-dataset/AFUN_eval))
are disjoint from this pool at the sample level.

## Citation

```bibtex
@article{wang2026afun,
  title   = {AFUN: Towards an Affordance Foundation Model for Functionality Understanding},
  author  = {Wang, Zhaoning and Zhong, Yi and Fu, Jiawei and Christensen, Henrik I. and Gao, Jun},
  journal = {arXiv preprint arXiv:2606.02551},
  year    = {2026}
}
```