---
license: cc-by-nc-4.0
task_categories:
- robotics
- image-to-3d
language:
- en
tags:
- egotouch
- tactile
- vitra
- mano
- hand-object-interaction
- corrected-mano
---

# egotouch-annotations-v1-leftfix

**Annotations only. This repository does not contain images or video.**

This release corrects the left-hand MANO rotation convention in
[`egotouch-annotations-v1`](https://huggingface.co/datasets/MIT-Media-Lab/egotouch-annotations-v1).
It keeps the original episode structure, instructions, tactile data, and sample index.

| Item | Count |
|---|---:|
| Episodes | 111,159 |
| Frames and training samples | 3,687,389 |
| Repaired left-hand episodes | 54,373 |
| Repaired left-hand frames | 1,781,845 |
| Unchanged right-only episodes | 56,786 |

## Correction

The source pipeline generated left-hand joints by mirroring a right-hand MANO model.
However, it stored the rotation matrices in the original right-hand convention.
This mismatch gave incorrect MANO-local coordinates for the left hand.

Let `M = diag(-1, 1, 1)`. This release applies these operations to left-hand episodes:

```python
global_orient_camspace = M @ global_orient_camspace @ M
hand_pose[:, joint] = M @ hand_pose[:, joint] @ M
global_orient_worldspace = R_camera_to_world @ global_orient_camspace
```

Reflection conjugation keeps each corrected matrix in `SO(3)`.
The correction does not mirror joints or translations again.

## Unchanged data

The following data is unchanged:

- `joints_camspace` and `joints_worldspace`
- `transl_camspace`, `transl_worldspace`, and wrist positions
- tactile pressure grids
- MANO shape parameters (`beta`)
- camera intrinsics and extrinsics
- RGB frame references (`video_decode_frame`)
- instructions and paraphrases
- episode membership and frame counts
- all right-hand annotations

## Files

```text
egotouch.tar
  Annotation/egotouch/episodic_annotations/*.npy

episode_frame_index.npz
  index_frame_pair       (3,687,389, 2) uint32
  index_to_episode_id    (111,159,)

full_repair_audit.json
  source and corrected hashes, counts, and validation results
```

The index is byte-identical to the parent release.
Each row of `index_frame_pair` gives an episode ordinal and a frame number.

## Load one sample

Extract `egotouch.tar`. Then load a sample with this code:

```python
import numpy as np

sample_id = 0
index = np.load("episode_frame_index.npz", allow_pickle=True)
episode_slot, frame_id = index["index_frame_pair"][sample_id]
episode_id = str(index["index_to_episode_id"][episode_slot])

path = f"Annotation/egotouch/episodic_annotations/{episode_id}.npy"
episode = np.load(path, allow_pickle=True).item()
rgb_frame_id = int(episode["video_decode_frame"][frame_id])
```

Each episode stores MANO parameters, camera data, world-space joints, camera-space joints,
tactile grids, instructions, paraphrases, and absolute RGB frame indices.
The `anno_type` field identifies the target hand for the episode.

## Coordinate convention

The per-frame `extrinsics` matrices use the world-to-camera convention.
The `global_orient_camspace` and `global_orient_worldspace` fields contain rotation matrices.
The world frame stays fixed within each source recording.

Chunk-level anchor-camera conversion belongs in the VITRA data loader.
This release does not bake an anchor-camera frame into the stored annotations.

## Validation

The full release passed these checks:

- all 111,159 archive members load successfully.
- archive names and order match the parent release.
- episode identifiers and frame counts match the index.
- no index row points outside an episode.
- the maximum rotation determinant error is `5.65e-7`.
- the maximum rotation orthogonality error is `6.34e-5`.
- the maximum camera-to-world orientation consistency error is `0.0`.
- sampled right-only episodes are byte-identical to the parent release.

The audit file contains the archive hashes and exact validation counts.

## Compatibility

Use this release for training that derives left-hand MANO-local coordinates from stored rotations.
Rebuild action caches that were made from the uncorrected left-hand rotation fields.

Caches that use only stored joints and tactile grids do not change numerically.
Their training protocol can still require a new comparison run for reproducibility.

## RGB frames

The `video_decode_frame` field indexes the source EgoTouch video.
Get the video from the [EgoTouch source release](https://huggingface.co/datasets/zhouzhoujy/EgoTouch).

## Integrity

```text
egotouch.tar
fe49162862d792861612c6405eebba958197a30df519a6e0fe1aa3e1db3f3088

episode_frame_index.npz
f9908d8733f00382a8fd762a1c8ab6f4b34fa47c5603c512c29fa66e31f47343
```