README.md

---
license: cc-by-nc-4.0
task_categories:
- robotics
language:
- en
tags:
- vitra
- mano
- hand-object-interaction
- tactile
- egocentric
---

# tachin-annotations-v1

**Annotations only — no images, no video.**

VITRA-style hand episodes for the **Tachin tactile glove** dataset, with per-hand instructions,
paraphrases, and **dense per-frame tactile**.

| | |
|---|---|
| episodes | **2,669** |
| training samples (`index_frame_pair` rows) | **79,584** |
| annotation | MANO pose + world/camera joints + per-frame extrinsics |
| **tactile** | **880 taxels per hand**, plus x/y shear, contact state, and 5 fingertip channels |
| text | one instruction per episode + 1.92 paraphrases on average |
| source frame rate | 30 fps |
| recordings | 102 task segments |
| **images / video** | **not included** — see *Getting the frames* below |

This is the **densest tactile** in the collection: 880 taxels per hand against EgoTouch's
16×16 = 256 per palm, and it additionally ships shear (`tactile_tf_x` / `tactile_tf_y`),
a per-taxel contact state, and 5 fingertip pressure and dynamics channels.

### What we did, and what we took from upstream

**The pose is not ours.** The upstream release
`MIT-Media-Lab/tachin-vitra-precut-mano-v1` did the hard part: the source ships 21-joint hand
tracks and **no MANO**, so MANO was fitted there (per-recording/hand `beta`, per-frame pose,
wrist `J0` translation, with a per-frame `fit_mpjpe_m`), the ~55.6 Hz glove streams were
interpolated onto the 30 Hz RGB timestamps without extrapolating across gaps wider than 100 ms,
and a 12.4-second disagreement between two published RGB clock origins was resolved with the
per-frame NTP table rather than a blanket offset. We adopted all of that unchanged.

**The episodes and the text are ours.** We re-cut the continuous recordings with VITRA's method:

    gaussian smooth (sigma=1.0) -> local speed minima in a fixed window (win=15, i.e. 0.5 s
    at 30 Hz) -> merge runs shorter than min_seg=16 -> pad 2 frames on each end

Left and right hands are cut independently. The validity mask is
`post_mano_cut_valid_frames`, the strictest of the four the source ships (median 0.955; the
others are `source_valid_frames` 0.968, `mano_fit_valid_frames` 0.967,
`projection_valid_frames` 0.967 — none is constant, so all four encode real validity).

**Instructions are ours.** Two rounds, both with `Qwen3.5-122B-A10B-FP8`: round 1 captions 8
frames per episode with the palm's future trajectory drawn on them; round 2 checks the sentence
belongs to that hand, strips same-hand references, and writes 1-3 paraphrases.

**Episodes with no instruction are not included** — round 1 returned `N/A` on 597 of 3,019
segments (19.8%); those are excluded from both the archive and the index.

Hands are balanced: 1,019 left / 1,403 right, i.e. left is 42% of episodes.

### Coordinates — no SLAM or anchor search needed

Per-frame `extrinsics` (World2Cam, OpenCV, metres) ship with every episode and one stable world
frame per recording. Measured: the camera really moves (max displacement median 0.196 m,
trajectory length 2–12 m), rotations are legal (orthogonality deviation ≤ 1.3e-07), and
`joints_camspace == extrinsics · joints_worldspace` to 8e-08 m. Anchor-camera normalisation is
**not** baked in; the training dataloader does chunk-level anchor synchronisation, the same as
every other domain here.

### Train / val / test split

`test` owns whole videos — its 5 videos appear in neither `train` nor `val`. `train` and `val`
share the remaining 97 videos and are separated at the episode level. Balanced on **frames**.

| split | videos | episodes | frames | share |
|---|---:|---:|---:|---:|
| train | 97 | 2,159 | 86,872 | 88.97% |
| val | 58 | 120 | 4,824 | 4.94% |
| test | 5 (exclusive) | 143 | 5,942 | 6.09% |

The test share overshoots 5% because there are only 102 videos and `test` must take whole ones;
the granularity is limited by video size, not by a bug.

### Files

```
tachin.tar                  ->  Annotation/tachin/episodic_annotations/*.npy
episode_frame_index.npz        index_frame_pair (N,2) uint32 + index_to_episode_id (E,)
splits/{train,val,test}.txt    episode ids
splits/test_videos.txt         the 5 videos test owns
splits/meta.json               parameters + achieved shares + self-checks
```

```python
import numpy as np
# tar -xf tachin.tar
z = np.load("episode_frame_index.npz", allow_pickle=True)
ep_slot, frame_id = z["index_frame_pair"][sample_id]
eid = str(z["index_to_episode_id"][ep_slot])
d = np.load(f"Annotation/tachin/episodic_annotations/{eid}.npy", allow_pickle=True).item()
rgb_frame_id = int(d["video_decode_frame"][frame_id])
hand = str(d["anno_type"])
tactile = d["tactile"][hand]          # (T, 880) float
valid   = d["tactile_valid"][hand]    # (T, 880) — NaNs are masked here, not filled
```

Each `.npy` also carries `tactile_tf_x` / `tactile_tf_y` (shear), `tactile_contact_state`,
`tactile_fingertip_pressure` and `tactile_fingertip_dynamics` (T,5), the static
`tactile_sensor_keys` (15) and `tactile_sensor_offsets` (15,2), a `quality` dict of per-frame
masks, and the usual `left`/`right` pose dicts.
`text[hand] = [(sentence, (0, T))]`, `text_rephrase[hand] = [([paraphrases...], (0, T))]`.

### Getting the frames

`video_decode_frame` indexes the **source** video, which we do not redistribute. The segment
videos are in `MIT-Media-Lab/tachin-vitra-precut-mano-v1` under
`Video/Tachin_root/<recording>/task-*.mp4`, and `video_name` in each episode is already the
full relative path from that repo root — join it directly, do not match by basename.
The original dataset is `Tachintech/TachinTactileGlove-Demo01`
(revision `2eb0e45e6d2103f33a7329c657a161c067d516c9`).

### Known limitations

- Paraphrase count averages 1.92, not a fixed number.
- **Small.** 2,422 episodes / 54 minutes of source video — the smallest domain in this
  collection. Its value is tactile density, not scale.
- MANO here is a **fit to 21-joint tracks**, not a multi-view solve; `fit_mpjpe_m` is stored
  per frame so the fit quality can be filtered on.
- Intrinsics are non-square (`fx` 626.9 vs `fy` 578.1) and imply (2cx, 2cy) = 1926.7 × 1069.3
  while the video is 1920 × 1080. Use the video's native size.
- Verified: the index lists exactly the episodes that have an instruction, every episode's
  stored frame count matches its index rows, and no index entry points at a missing episode.

---
## Revision — 2026-09-08 (re-cut)

**This release replaces the previous one. The previous episodes contained invalid frames
and should not be used.**

A defect in our episode-cutting step let frames with `kept_frames == False`
(invalid hand pose — all-zero or NaN wrist coordinates) stay inside published episodes.
The validity mask was only used to keep a cut point from landing on an invalid frame; it did
not constrain what a segment contained. Worse, a run of invalid frames could suppress cutting
altogether, so the gap was swallowed into one long segment instead of being excluded.

Every episode here is now built from a run of consecutive valid frames, so `kept_frames` is
all-`True` by construction — verified over the whole collection: **568,369 episodes /
16.5 M frames, zero `kept_frames == False`**. Both instruction rounds were regenerated for the
new segmentation.

Segment counts and episode ids therefore changed, and the index and splits were rebuilt:

| | previous | this release |
|---|---:|---:|
| episodes on disk | 3,019 | 3,420 |
| episodes published (with an instruction) | 2,422 | 2,669 |
| training samples | 97,638 | 79,584 |

Splits are `video_test` (test owns whole videos disjoint from train/val; train and val share
the remaining videos and are split at the episode level), balanced on **frames** at 90/5/5,
seed 1.

<!-- COLLECTION_TABLE_START -->
| dataset | episodes | training samples | our contribution | size | HF |
|---|---:|---:|---|---:|---|
| **EPIC-KITCHENS-100** | 149,570 | 4,019,534 | episodes + text | 8.70 GB | [`epic30-annotations-v1`](https://huggingface.co/datasets/MIT-Media-Lab/epic30-annotations-v1) |
| **EgoTouch** | 107,364 | 3,123,675 | episodes + text + **tactile** | 17.02 GB | [`egotouch-annotations-v1`](https://huggingface.co/datasets/MIT-Media-Lab/egotouch-annotations-v1) |
| **GigaHands** | 70,486 | 2,266,087 | episodes + text | 2.89 GB | [`gigahands-annotations-v1`](https://huggingface.co/datasets/MIT-Media-Lab/gigahands-annotations-v1) |
| **Ego-Exo4D** | 67,051 | 1,757,474 | text only | 4.09 GB | [`egoexo4d-annotations-v1`](https://huggingface.co/datasets/MIT-Media-Lab/egoexo4d-annotations-v1) |
| **OakInk2** | 29,058 | 1,052,924 | episodes + text | 1.56 GB | [`oakink2-annotations-v1`](https://huggingface.co/datasets/MIT-Media-Lab/oakink2-annotations-v1) |
| **TACO** | 23,757 | 736,136 | episodes + text | 1.34 GB | [`taco-annotations-v1`](https://huggingface.co/datasets/MIT-Media-Lab/taco-annotations-v1) |
| **HOT3D** | 18,805 | 619,680 | episodes + text | 1.51 GB | [`hot3d-annotations-v1`](https://huggingface.co/datasets/MIT-Media-Lab/hot3d-annotations-v1) |
| **ARCTIC** | 12,610 | 425,796 | episodes + text | 0.85 GB | [`arctic-annotations-v1`](https://huggingface.co/datasets/MIT-Media-Lab/arctic-annotations-v1) |
| **H2O** | 5,845 | 200,332 | episodes + text | 0.40 GB | [`h2o-annotations-v1`](https://huggingface.co/datasets/MIT-Media-Lab/h2o-annotations-v1) |
| **Tachin** | 2,669 | 79,584 | episodes + text + **tactile** | 3.38 GB | [`tachin-annotations-v1`](https://huggingface.co/datasets/MIT-Media-Lab/tachin-annotations-v1) |
| **total** | **487,215** | **14,281,222** | | **41.7 GB** | |

> **Something-Something V2 was dropped** from the collection (12 fps against
> 30 fps everywhere else, so a 16-step action chunk spans 1.33 s instead of 0.53 s).
> The repository still exists but should not be used.
<!-- COLLECTION_TABLE_END -->