dataset card

README.md

@@ -0,0 +1,139 @@
+# handTrackingSample
+
+Egocentric footage of people working with their hands, cut to the sections a human picked as
+hand-tracked, with synchronised IMU and per-frame hand pose.
+
+**149 episodes · 10.6 hours · 46 recordings · 27 cameras · 1920x1200 · stereo**
+
+---
+
+## What each episode contains
+
+```
+<run>/<seg>/
+  left.mp4      1920x1200 H.265, white-balanced, lens-rectified, faces pixelated
+  right.mp4     the other eye of the stereo pair, same treatment
+  frames.csv    frame_idx, ts_us      <- the authoritative video clock
+  imu.csv       ts_us, ax/ay/az (m/s2), gx/gy/gz (deg/s)
+  hands.jsonl   per-frame hand + held-object boxes, and 21-joint hand pose
+  labels.json   scene, task, job, environment, zone provenance
+  _stat.json    processing record, including the exact anonymisation setting
+```
+
+## How the sections were chosen
+
+A person reviewed the recordings and drew the hand-tracked ranges by hand. Those ranges -- not a
+detector -- define the episodes here. An independent hand detector agrees with them: every curated
+zone has hand detections in at least 52.9% of its frames, median 92.3%.
+
+---
+
+## Read this before you align anything
+
+**Use `frames.csv`, never the container frame rate.** The mp4 carries a synthetic constant-rate
+PTS. The true exposure clock differs from it by up to **135 ms across an episode** -- about four
+frames. Anyone computing frame time as `index / fps` will drift out of sync with the IMU by the
+end. `frames.csv` carries the real per-frame timestamp, on the same clock as `imu.csv`.
+
+**`dt` is not uniform.** Recordings contain genuine dropped-frame gaps, up to about **0.5 s**.
+Video and IMU stay locked to each other across them -- the sample-per-frame ratio is constant -- but
+integrating gyro on an assumed fixed `dt` will accumulate error. Gaps are listed per episode in
+`_stat.json`.
+
+**`hands.jsonl` is sparse and keyed by `frame_idx`.** Not every frame has a row, and pose is
+computed on a stride. Index by `frame_idx`; positional indexing will silently misalign.
+
+---
+
+## Anonymisation -- known limitation
+
+Faces are detected with SCRFD-10G at a 960 letterbox, upright threshold 0.50, unioned with a
+four-rotation pass at 0.65, and pixelated with an oval mask.
+
+That setting was chosen on a **hand-labelled benchmark of 31 faces drawn from this same footage**,
+where it catches **68% of them at a 2.7% false-positive rate**. The rotation pass is what makes it
+work: head-mounted video puts faces at in-plane angles an upright detector scores near zero on --
+two benchmark faces score 0.07 and 0.08 upright and 0.70 and 0.76 rotated.
+
+**68% is an upper bound on what was caught, not a floor on what leaked.** The benchmark can only
+contain faces some detector saw at all, so faces no model detected are absent from both the
+numerator and the denominator. **Identifiable faces remain in this data.** Do not use it for face
+recognition, re-identification, or any purpose requiring that individuals be unidentifiable.
+
+## Rectification
+
+Each eye is undistorted with **its own camera's calibration** (intrinsics + 8-parameter rational
+distortion), not a shared one -- the units are the same 116-degree optics but their coefficients
+differ enough that a shared correction would be wrong.
+
+Calibrations report stereo RMSE around 0.2-0.7 px and pass a live 128-sample check. The correction
+is zero at the optical centre and grows with radius, reaching roughly 12-96 px at the frame edge
+depending on the unit. One caveat: RMSE only measures fit where the calibration target sat, so
+far-corner geometry is extrapolated rather than verified.
+
+Only cameras with a calibration are included here. Every camera excluded for lacking one is on the
+project's known-faulty list.
+
+## Hand pose
+
+21 3D joints per hand from **WiLoR** (ViT detector + MANO mesh head), run every 6 frames.
+MediaPipe's landmarker was evaluated against WiLoR and EgoHOS on this footage and was not good
+enough -- it is the same model family that came last of eight detectors on our face benchmark, and
+head-mounted video is its weak case.
+
+```json
+{"frame_idx": 120,
+ "pose": [{"kp2d": [[x,y], ...21], "kp3d": [[x,y,z], ...21],
+           "is_right": true, "box": [x1,y1,x2,y2]}]}
+```
+
+`kp2d` is in frame pixels; `box` is WiLoR's own hand detection, so detection and articulation come
+from the same model and always agree. `kp3d` is WiLoR's 3D joint prediction in its own metric
+frame, not camera millimetres.
+
+**Held-object boxes are not in this release.** The 100DOH pass that labels what a hand is *holding*
+runs separately and was not applied here; `hands.jsonl` carries hand detection and pose only.
+
+## Scene and task spread
+
+149 episodes drawn from 46 recordings on 27 different cameras, across **13 top-level categories**:
+
+| category | recordings |
+|---|---|
+| Creative Workshops | 8 |
+| Repair Services | 6 |
+| Printing and Design | 5 |
+| Food and Beverage | 5 |
+| Retail and Consumer Goods | 5 |
+| Cleaning and Sanitation | 4 |
+| Laboratory / Scientific | 3 |
+| Automotive and Transport | 3 |
+| Clothing and Fashion | 2 |
+| Animal Care | 2 |
+| Energy and Utilities | 1 |
+| Industrial Manufacturing | 1 |
+| Agriculture and Farming | 1 |
+
+No category exceeds 8 of 46 recordings. Episodes were selected round-robin across cameras, so the
+set spans the full camera fleet rather than over-sampling whichever recordings happened to be
+longest.
+
+## How much of the footage carries a hand skeleton
+
+Pose is computed every 6th frame. The share of those sampled frames that carry at least one hand
+varies a lot by task, and is a property of the work, not a failure of the pipeline:
+
+- typical episodes land around **60-95%**
+- bench work with both hands in view sits at the top of that range
+- **overhead work is the weak case** -- an electrician working at ceiling height gave 1.8%, because
+  the camera spends most of the episode pointed at bare wall, and when the hands do appear the
+  wearer is reaching UP, so the arms enter from the top of frame. WiLoR is trained mostly on hands
+  entering from below, and an inverted approach is its hard case.
+
+Check `hands.jsonl` before assuming a given episode is densely annotated.
+
+## Provenance
+
+Every episode records its source recording, the exact source frame range, the curated zone it came
+from, and the anonymisation setting used, in `_stat.json`. Cut and rendered from raw with
+`generateSample/run_sample.py`.