hishamstarpilot commited on
Commit
e4ae10b
·
verified ·
1 Parent(s): f546b05

dataset card

Browse files
Files changed (1) hide show
  1. README.md +53 -40
README.md CHANGED
@@ -1,17 +1,20 @@
1
  # handTrackingSample
2
 
3
  Egocentric footage of people working with their hands, cut to the sections a human picked as
4
- hand-tracked, with synchronised IMU and per-frame hand pose.
5
 
6
- **149 episodes · 10.6 hours · 46 recordings · 27 cameras · 1920x1200 · stereo**
7
 
8
  ---
9
 
10
- > **18 Aug 2026 — partial takedown, re-render in progress.** 64 of the original 149 episodes have
11
- > been temporarily removed: they were rectified with a per-camera calibration whose undistortion
12
- > map folds inside the frame, producing a visible circular seam. All 149 are being re-rendered
13
- > without rectification and with a stricter face-blur setting, and will replace this set shortly.
14
- > The 85 episodes currently here are unaffected by the seam.
 
 
 
15
 
16
 
17
  ## What each episode contains
@@ -22,7 +25,7 @@ hand-tracked, with synchronised IMU and per-frame hand pose.
22
  right.mp4 the other eye of the stereo pair, same treatment
23
  frames.csv frame_idx, ts_us <- the authoritative video clock
24
  imu.csv ts_us, ax/ay/az (m/s2), gx/gy/gz (deg/s)
25
- hands.jsonl per-frame hand + held-object boxes, and 21-joint hand pose
26
  labels.json scene, task, job, environment, zone provenance
27
  _stat.json processing record, including the exact anonymisation setting
28
  ```
@@ -47,25 +50,31 @@ Video and IMU stay locked to each other across them -- the sample-per-frame rati
47
  integrating gyro on an assumed fixed `dt` will accumulate error. Gaps are listed per episode in
48
  `_stat.json`.
49
 
50
- **`hands.jsonl` is sparse and keyed by `frame_idx`.** Not every frame has a row, and pose is
51
- computed on a stride. Index by `frame_idx`; positional indexing will silently misalign.
 
52
 
53
  ---
54
 
55
  ## Anonymisation -- known limitation
56
 
57
- Faces are detected with SCRFD-10G at a 960 letterbox, upright threshold 0.50, unioned with a
58
- four-rotation pass at 0.65, and pixelated with an oval mask.
 
59
 
60
- That setting was chosen on a **hand-labelled benchmark of 31 faces drawn from this same footage**,
61
- where it catches **68% of them at a 2.7% false-positive rate**. The rotation pass is what makes it
62
- work: head-mounted video puts faces at in-plane angles an upright detector scores near zero on --
63
- two benchmark faces score 0.07 and 0.08 upright and 0.70 and 0.76 rotated.
 
 
64
 
65
- **68% is an upper bound on what was caught, not a floor on what leaked.** The benchmark can only
66
  contain faces some detector saw at all, so faces no model detected are absent from both the
67
- numerator and the denominator. **Identifiable faces remain in this data.** Do not use it for face
68
- recognition, re-identification, or any purpose requiring that individuals be unidentifiable.
 
 
69
 
70
  ## Lens geometry
71
 
@@ -77,7 +86,7 @@ raw geometry was judged the better trade. Per-camera intrinsics are available on
77
 
78
  ## Hand pose
79
 
80
- 21 3D joints per hand from **WiLoR** (ViT detector + MANO mesh head), run every 6 frames.
81
  MediaPipe's landmarker was evaluated against WiLoR and EgoHOS on this footage and was not good
82
  enough -- it is the same model family that came last of eight detectors on our face benchmark, and
83
  head-mounted video is its weak case.
@@ -85,27 +94,38 @@ head-mounted video is its weak case.
85
  ```json
86
  {"frame_idx": 120,
87
  "pose": [{"kp2d": [[x,y], ...21], "kp3d": [[x,y,z], ...21],
88
- "is_right": true, "box": [x1,y1,x2,y2]}]}
 
89
  ```
90
 
91
- `kp2d` is in frame pixels; `box` is WiLoR's own hand detection, so detection and articulation come
92
- from the same model and always agree. `kp3d` is WiLoR's 3D joint prediction in its own metric
93
- frame, not camera millimetres.
 
 
 
 
 
 
94
 
95
- **Held-object boxes are not in this release.** The 100DOH pass that labels what a hand is *holding*
96
- runs separately and was not applied here; `hands.jsonl` carries hand detection and pose only.
 
 
 
 
97
 
98
  ## Scene and task spread
99
 
100
- 149 episodes drawn from 46 recordings on 27 different cameras, across **13 top-level categories**:
101
 
102
  | category | recordings |
103
  |---|---|
104
  | Creative Workshops | 8 |
105
  | Repair Services | 6 |
106
  | Printing and Design | 5 |
107
- | Food and Beverage | 5 |
108
- | Retail and Consumer Goods | 5 |
109
  | Cleaning and Sanitation | 4 |
110
  | Laboratory / Scientific | 3 |
111
  | Automotive and Transport | 3 |
@@ -115,21 +135,14 @@ runs separately and was not applied here; `hands.jsonl` carries hand detection a
115
  | Industrial Manufacturing | 1 |
116
  | Agriculture and Farming | 1 |
117
 
118
- No category exceeds 8 of 46 recordings. Episodes were selected round-robin across cameras, so the
119
  set spans the full camera fleet rather than over-sampling whichever recordings happened to be
120
  longest.
121
 
122
  ## How much of the footage carries a hand skeleton
123
 
124
- Pose is computed every 6th frame. The share of those sampled frames that carry at least one hand
125
- varies a lot by task, and is a property of the work, not a failure of the pipeline:
126
-
127
- - typical episodes land around **60-95%**
128
- - bench work with both hands in view sits at the top of that range
129
- - **overhead work is the weak case** -- an electrician working at ceiling height gave 1.8%, because
130
- the camera spends most of the episode pointed at bare wall, and when the hands do appear the
131
- wearer is reaching UP, so the arms enter from the top of frame. WiLoR is trained mostly on hands
132
- entering from below, and an inverted approach is its hard case.
133
 
134
  Check `hands.jsonl` before assuming a given episode is densely annotated.
135
 
@@ -137,4 +150,4 @@ Check `hands.jsonl` before assuming a given episode is densely annotated.
137
 
138
  Every episode records its source recording, the exact source frame range, the curated zone it came
139
  from, and the anonymisation setting used, in `_stat.json`. Cut and rendered from raw with
140
- `generateSample/run_sample.py`.
 
1
  # handTrackingSample
2
 
3
  Egocentric footage of people working with their hands, cut to the sections a human picked as
4
+ hand-tracked, with synchronised IMU and hand pose on **every frame**.
5
 
6
+ **120 episodes · 8.4 hours · 44 recordings · 27 cameras · 1920x1200 · stereo**
7
 
8
  ---
9
 
10
+ > **19 Aug 2026 — RE-UPLOAD IN PROGRESS.** The 120 re-rendered episodes are being pushed now and
11
+ > will appear here over the next hour or so; the file tree below is incomplete until then.
12
+ >
13
+ > **Full re-release.** Every episode here was re-rendered from raw: no lens
14
+ > rectification (the earlier release had a visible circular seam on cameras whose undistortion map
15
+ > folds inside the frame), a stricter face-blur setting (see *Anonymisation*), no orientation flip
16
+ > guessing, and WiLoR hand pose on every frame rather than every sixth. The earlier 149-episode
17
+ > set has been withdrawn in full; this set replaces it.
18
 
19
 
20
  ## What each episode contains
 
25
  right.mp4 the other eye of the stereo pair, same treatment
26
  frames.csv frame_idx, ts_us <- the authoritative video clock
27
  imu.csv ts_us, ax/ay/az (m/s2), gx/gy/gz (deg/s)
28
+ hands.jsonl per-frame hand boxes, 21-joint 2D/3D pose and MANO parameters
29
  labels.json scene, task, job, environment, zone provenance
30
  _stat.json processing record, including the exact anonymisation setting
31
  ```
 
50
  integrating gyro on an assumed fixed `dt` will accumulate error. Gaps are listed per episode in
51
  `_stat.json`.
52
 
53
+ **`hands.jsonl` is keyed by `frame_idx`.** There is one row per decoded frame, but index by
54
+ `frame_idx` rather than by position -- it is the same index `frames.csv` uses, and it is what ties
55
+ a row to its timestamp.
56
 
57
  ---
58
 
59
  ## Anonymisation -- known limitation
60
 
61
+ Faces are detected with SCRFD-10G at a 960 letterbox, upright threshold **0.80**, unioned with a
62
+ four-rotation pass at **0.80**, boxes over 10% of the frame additionally required to score 0.80,
63
+ and pixelated with an oval mask held for 6 frames after the last detection.
64
 
65
+ This is a deliberately **conservative** setting, chosen after the earlier release pixelated hands,
66
+ cookware and packaging that the detector scored as faces. On a hand-labelled benchmark of 31 faces
67
+ drawn from this same footage it catches **5 of 31 (16%)** with one false positive in 140 frames;
68
+ the earlier release's setting caught 68% with eight. The faces it does catch are the frontal,
69
+ close ones -- a bystander in profile or a face at an angle to the camera will usually score below
70
+ 0.80 and pass through unblurred.
71
 
72
+ **16% is an upper bound on what was caught, not a floor on what leaked.** The benchmark can only
73
  contain faces some detector saw at all, so faces no model detected are absent from both the
74
+ numerator and the denominator. **Identifiable faces remain in this data, more than in most
75
+ published egocentric sets.** Do not use it for face recognition, re-identification, or any purpose
76
+ requiring that individuals be unidentifiable. `_stat.json` records exactly what was applied, so
77
+ a stronger pass can be layered on top with your own detector.
78
 
79
  ## Lens geometry
80
 
 
86
 
87
  ## Hand pose
88
 
89
+ 21 3D joints per hand from **WiLoR** (ViT backbone + MANO mesh head), on **every frame**.
90
  MediaPipe's landmarker was evaluated against WiLoR and EgoHOS on this footage and was not good
91
  enough -- it is the same model family that came last of eight detectors on our face benchmark, and
92
  head-mounted video is its weak case.
 
94
  ```json
95
  {"frame_idx": 120,
96
  "pose": [{"kp2d": [[x,y], ...21], "kp3d": [[x,y,z], ...21],
97
+ "is_right": true, "box": [x1,y1,x2,y2],
98
+ "mano": {"global_orient": [3], "hand_pose": [45], "betas": [10], "cam_t": [3]}}]}
99
  ```
100
 
101
+ - `kp2d` is in frame pixels; `box` is WiLoR's own hand detection, so detection and articulation
102
+ come from the same model and always agree.
103
+ - `kp3d` is WiLoR's 3D joint prediction in its own metric frame (metres, hand-rooted), not camera
104
+ millimetres. `mano` carries the axis-angle global orientation, the 15 joint rotations, the shape
105
+ coefficients and the weak-perspective camera translation, so the full mesh can be rebuilt with a
106
+ standard MANO layer.
107
+ - Left hands are predicted by mirroring, as in WiLoR; `kp3d.x`, `global_orient` and `hand_pose`
108
+ are already mirrored back into the un-flipped frame.
109
+ - Frames with no detected hand carry `"pose": []`.
110
 
111
+ Pose was computed with a batched re-implementation of WiLoR-mini's inference loop (crop-local
112
+ anti-alias blur instead of a full-frame blur per hand). Against the stock pipeline on 102 hands it
113
+ gives identical detections and a median 1.0 px / max 7.6 px difference in `kp2d`, which is the
114
+ model's own sensitivity to sub-grey-level input rounding.
115
+
116
+ **Held-object boxes are not in this release.** `hands.jsonl` carries hand detection and pose only.
117
 
118
  ## Scene and task spread
119
 
120
+ 120 episodes drawn from 44 recordings on 27 different cameras, across **13 top-level categories**:
121
 
122
  | category | recordings |
123
  |---|---|
124
  | Creative Workshops | 8 |
125
  | Repair Services | 6 |
126
  | Printing and Design | 5 |
127
+ | Food and Beverage | 4 |
128
+ | Retail and Consumer Goods | 4 |
129
  | Cleaning and Sanitation | 4 |
130
  | Laboratory / Scientific | 3 |
131
  | Automotive and Transport | 3 |
 
135
  | Industrial Manufacturing | 1 |
136
  | Agriculture and Farming | 1 |
137
 
138
+ No category exceeds 8 of 44 recordings. Episodes were selected round-robin across cameras, so the
139
  set spans the full camera fleet rather than over-sampling whichever recordings happened to be
140
  longest.
141
 
142
  ## How much of the footage carries a hand skeleton
143
 
144
+ The share of frames that carry at least one hand varies a lot by task, and is a property of the
145
+ work, not a failure of the pipeline: (per-episode figures will be filled in when the upload completes).
 
 
 
 
 
 
 
146
 
147
  Check `hands.jsonl` before assuming a given episode is densely annotated.
148
 
 
150
 
151
  Every episode records its source recording, the exact source frame range, the curated zone it came
152
  from, and the anonymisation setting used, in `_stat.json`. Cut and rendered from raw with
153
+ `generateSample/run_sample.py`; pose with `generateSample/wilor_pose.py`.