--- license: other task_categories: - robotics tags: - hand-pose - mano - egocentric - vitra - gigahands --- # GigaHands → VITRA Stage-1, official-MANO annotations VITRA Stage-1 hand annotations for **GigaHands**, with **all joint positions taken from GigaHands' official MANO fit** instead of mixing in triangulated keypoints. **Annotations only — no videos** (get those from GigaHands; the mapping is described in §5). | | | |---|---| | episodes | **13,247** (train 11,904 / test 1,343) | | frames | 3,395,733 | | camera | `brics-odroid-001_cam0` (static rig; one constant extrinsic per scene) | | source | GigaHands `params/` + `keypoints_3d_mano/` + `optim_params.txt` | --- ## 1. Three differences from the usual conversion VITRA's converter by default takes **joints from `keypoints_3d`** (multi-view triangulation) while taking **wrist 6DoF from `params`** (the MANO fit). This dataset takes **everything from the official MANO fit**. Three changes, each with measurements. ### (a) `joints_worldspace`: triangulation → official MANO Joints are taken directly from the official `keypoints_3d_mano`; **not a single coordinate is recomputed**. That file has its per-frame centroid zeroed out (measured: per-frame centroid std = 0), so only the translation is missing. It is restored as: ``` wrist_world(t) = R(Rh_t) · j0 + Th_t # official params + the constant j0 only c(t) = wrist_world(t) − wrist in the official points(t) joints_world(t) = official points(t) + c(t) ``` `j0` is the model-space wrist. It depends only on `beta`, not on finger pose (measured: per-frame std = 0, ‖j0‖ ≈ 0.096 m), so it is computed once per sequence. **This is the only quantity we compute ourselves.** > The official file zeroes the centroid over **both hands' 42 points jointly**, so `c(t)` is shared > between hands. This conversion solves it from the **right** hand and applies it to both — solving it > independently for the left hand requires approximating its `j0` with a mirrored right-hand model, > which we measured to deviate by **1.75–2.96 cm (max 11.6 cm)** from an independent pipeline, versus > **0.25–0.29 cm** for the right hand. **Why the change** — hand-size stability of the two annotation sources (sum of 20 bone lengths, right hand, first frame, metres): | scene | official MANO | triangulated | |---|---|---| | p001-packing | 0.772 | 1.104 | | p002-dog | 0.783 | 0.995 | | p002-firstaid | 0.786 | 0.971 | | p002-packing | 0.774 | 0.797 | | p003-instrument | 0.784 | 0.869 | | p003-packing | 0.777 | 0.905 | | **spread** | **1.8%** | **38.5%** | `p002-dog` (0.995) and `p002-packing` (0.797) are the **same participant**; the triangulated side differs by **25%** between them. ### (b) `transl_worldspace`: `Th` → wrist We store `joints_worldspace[:, 0]` (the wrist), not MANO's `Th`. The two differ by a constant **9.6 cm**. Rationale: VITRA's other domains store the wrist (`data/tools/hand_recon_core.py:118`, `transl_new = wrist + transl`). In this dataset `transl_worldspace == joints_worldspace[:,0]` holds **exactly, frame by frame** (sampled max difference 0.0e+00). ### (c) `hand_pose`: add `hands_mean` We store `matrix(poses[3:] + hands_mean)`. GigaHands' `poses` are relative to the MANO **mean hand**, whereas VITRA's other domains store HaWoR's **absolute** rotation matrices (`MANOLayer` runs with `pose2rot=False`, so the `full_pose += pose_mean` at `smplx/body_models.py:184` never executes). Adding it back puts both on the same convention. **Test** — how well the result reproduces the official `keypoints_3d_mano` (6 scenes): | variant | mean error | |---|---| | without (`flat_hand_mean=True`) | 3.39 cm | | **with `hands_mean`** | **1.29 cm** | > ⚠ The GigaHands paper states "PCA disabled + **flat mean** shape", which **disagrees** with this > measurement. This dataset follows the measurement. ### Global-transform convention (determined by measurement, not assumed) Using `keypoints_3d` as a yardstick, one sequence from each of 6 scenes, mean per-joint Euclidean distance: | convention | left hand | left error | right error | |---|---|---|---| | **`v_world = R(Rh)·v_model + Th`** (about the model origin) | **mirrored** | **6.91 cm** | **6.71 cm** | | `v_world = R·(v−j0) + j0 + Th` (about the root joint) | mirrored | 16.41 cm | 15.52 cm | | about the model origin | not mirrored | 10.32 cm | 6.71 cm | | about the root joint | not mirrored | 18.34 cm | 15.52 cm | The left hand uses the **right-hand model with the x-axis flipped**, matching VITRA (`hand_recon_core.py:112-115`); `MANO_LEFT.pkl` is never loaded. --- ## 2. ⚠ Known issue: 453 episodes we believe are image↔label misaligned (**please verify independently**) `residual_misaligned_episodes.json` lists **453 / 13,247 (3.42%)** episodes whose annotation length differs from the frame count of the video they are bound to by **more than 12 frames**: ``` GigaHands_p001-packing_000_... params 381 video 483 mano 381 GigaHands_p001-packing_043_... params 583 video 702 mano 583 GigaHands_p002-packing_002_... params 177 video 336 mano 177 ``` The pattern is consistent: **`params` and `keypoints_3d_mano` always agree with each other, and differ from the video by 15–159 frames** — not the small difference you would expect from tail trimming. **Our reading** is that the annotated time span does not correspond to the video, so the hand at frame *t* and the image at frame *t* are not the same instant. **This is only our reading. There may be an alignment scheme we have not considered (some timestamp mapping, for instance). These episodes are still included in the dataset — please check them yourself and decide whether to drop them. If you find that this is not actually a problem, or you know the correct alignment, we would very much like to hear it.** For contrast: episodes where `params` length ≠ `keypoints_3d_mano` length: **0 / 13,247 (0.00%)**. ### A second, unquantified problem A few episodes have **all three lengths agreeing, yet projection shows the annotation does not match the video content**. Example: `p001-packing/023` (`params 225 / keypoints 225 / video 232`) — projecting its annotation onto that video puts the skeleton nowhere near the hands; shifting the video index by ±2 does not help; projecting the triangulated keypoints instead fails identically. **This class cannot be detected by length checks. Its size is unknown and it has not been removed.** --- ## 3. What was excluded | criterion | count | removed? | |---|---|---| | `low_valid_ratio < 0.9` (too few valid triangulated frames) | 209 | ✅ | | missing `keypoints_3d_mano` (`p017-fastfood/049`) | 1 | ✅ | | clip too short (`min_frames=32`) | 1 | ✅ | | the 453 suspected misalignments above | 453 | ❌ **kept** — see §2 | **About those 209**: the new pipeline uses MANO-fit output, whose values are always finite, so the `low_valid_ratio` check **always passes** under it (our raw run reported `skipped_low_valid_ratio = 0`). **That is not the data getting better — it is the check going blind.** Frames where triangulation failed get filled in by the MANO fit with plausible-looking poses, which is harder to notice. We therefore reused the same exclusion list derived from triangulation validity. --- ## 4. Contents and format ``` Annotation/ ├── gigahands_real_train/ │ ├── episodic_annotations.tar.zst # 11,904 .npy files │ ├── episode_frame_index.npz # 3,127,323 (episode_idx, frame_idx) pairs │ └── conversion_report.json ├── gigahands_real_test/ │ ├── episodic_annotations.tar.zst # 1,343 .npy files │ ├── episode_frame_index.npz │ └── conversion_report.json └── statistics/ ├── gigahands_real_train_keypoints_statistics.json └── gigahands_real_test_keypoints_statistics.json subset_manifest.json # 13,249 clips (includes the 2 that were dropped) residual_misaligned_episodes.json # the 453 from §2 ``` Each `.npy` is a `dict`; `left` and `right` each contain: | field | shape | notes | |---|---|---| | `beta` | (10,) | MANO shape, from the official `shapes` | | `global_orient_worldspace` | (T,3,3) | `matrix(Rh)` | | `global_orient_camspace` | (T,3,3) | left-multiplied by `R_w2c` | | `hand_pose` | (T,15,3,3) | `matrix(poses[3:] + hands_mean)` | | `transl_worldspace` | (T,3) | **wrist**, `= joints_worldspace[:,0]` | | `transl_camspace` | (T,3) | | | `joints_worldspace` | (T,21,3) | official MANO joints + restored translation, OpenPose order (wrist = 0) | | `joints_camspace` | (T,21,3) | | | `kept_frames` | (T,) bool | | Also present: `text` / `text_rephrase` (language instruction), `video_name`, `video_decode_frame`, `intrinsics`, `extrinsics`, `camera`. **16-frame windows are cut at training time; they are not the storage granularity.** Each `.npy` holds a whole segment (length 32–8996, median 175); `episode_frame_index.npz` enumerates every frame as a possible window start. --- ## 5. How this maps to the videos (videos are not in this repo) Get the videos from **GigaHands**. The episode filename encodes everything needed to locate them: ``` GigaHands_____f__ep_000000.npy │ │ │ │ │ └─ end frame in the source video (exclusive) │ │ │ │ └─ start frame │ │ │ └─ video filename without .mp4 │ │ └─ brics-odroid-001_cam0 │ └─ GigaHands sequence id └─ GigaHands scene name ``` The video lives at `multiview_rgb_vids///.mp4`. Each clip in `subset_manifest.json` also records `video_path` / `start_frame` / `end_frame` directly. > ⚠ **Sequence ids do not correspond one-to-one with the sorted position of video files in a > directory** — sequence names have gaps (e.g. `p003-instrument` has no `023`, so `'024'` sits at > sorted index 23). **Locate videos by the `video_stem` in the filename, never by index.** Camera intrinsics and extrinsics are in each `.npy` under `intrinsics` / `extrinsics`, taken from the `brics-odroid-001_cam0` row of GigaHands' official `optim_params.txt`. **This is a static rig: one constant extrinsic per scene** (the converter replicates it across frames), so no SLAM is needed and no per-frame alignment is required within a window. --- ## 6. What was not done (please do not infer quality from this) - **No A/B training comparison was run.** "Official-MANO is better than triangulated" is currently supported only by **indirect evidence** (bone-length spread 1.8% vs 38.5%; normalization std about 33× smaller) — **there is no training result behind it.** - **The normalizer changed accordingly.** The `statistics` here are **not interchangeable** with those of a triangulated-joint version; MSE values across different normalizers are **not directly comparable** and must be de-normalized to MANO physical units first. - The second problem in §2 (all lengths agree but content does not) is **unquantified and not removed**. --- ## Citation and licence This repository contains **derived annotations of GigaHands**. The original data, videos and MANO fit parameters remain the property of the GigaHands authors; please follow the GigaHands licence and cite their paper. The format follows **VITRA** (microsoft/VITRA) Stage-1. The MANO model itself is under the MPI licence and must be obtained from https://mano.is.tue.mpg.de — no MANO model file is included here.