update: complete dataset card
Browse files
README.md
CHANGED
|
@@ -61,15 +61,6 @@ ChingMu 1000H is an optical motion capture dataset designed for training and val
|
|
| 61 |
|
| 62 |
---
|
| 63 |
|
| 64 |
-
## β¨ Data Highlights
|
| 65 |
-
|
| 66 |
-
- **Optical Ground Truth** β Sub-millimeter accuracy (<1mm), 120 fps, no estimation errors.
|
| 67 |
-
- **Dexterous Hands** β 20+ DoF per hand, synchronized with object 6D pose for fine-grained manipulation.
|
| 68 |
-
- **Robot-Ready** β Pre-retargeted to Unitree G1; custom retargeting available on request.
|
| 69 |
-
- **Real-World Diversity** β 15+ scenarios (industrial, household, retail, healthcare, logistics, agriculture, performance), 500+ tasks, 1000+ objects.
|
| 70 |
-
- **Multi-Modal** β Full-body skeleton (BVH), finger motion, object pose, multi-view video, semantic labels.
|
| 71 |
-
- **Quality Assured** β Every take passes automated cleaning + manual inspection; quality flags (pass/warning/fail) provided.
|
| 72 |
-
|
| 73 |
## π Data Format Specifications
|
| 74 |
|
| 75 |
| Component | Format | Details |
|
|
@@ -79,20 +70,6 @@ ChingMu 1000H is an optical motion capture dataset designed for training and val
|
|
| 79 |
| Object 6D pose | `.csv` | Position (m) + quaternion per object, 120 Hz |
|
| 80 |
| Multi-view video | `.mp4` | 2β4 synchronized cameras, co-registered timeline |
|
| 81 |
|
| 82 |
-
## β
Quality Assurance
|
| 83 |
-
|
| 84 |
-
All data undergoes a rigorous pipeline:
|
| 85 |
-
1. **Marker swap correction** β automatic detection and repair.
|
| 86 |
-
2. **Gap filling** β cubic spline interpolation for gaps β€6 frames; longer gaps flagged.
|
| 87 |
-
3. **Foot skating detection** β flagged when static foot drifts >2cm/s.
|
| 88 |
-
4. **Manual review** β each take reviewed by trained annotators.
|
| 89 |
-
5. **Quality flags** β `pass` (clean), `warning` (minor issues), `fail` (do not use).
|
| 90 |
-
|
| 91 |
-
Typical accuracy:
|
| 92 |
-
- Joint position error: <1mm (optical system limit)
|
| 93 |
-
- Object 6D pose: Β±2mm translation, Β±0.5Β° rotation
|
| 94 |
-
- Temporal sync between modalities: <1 frame
|
| 95 |
-
|
| 96 |
## πΈ Sample Visualization
|
| 97 |
|
| 98 |
Display is a snapshot from our motion capture studio showing a subject performing a box-moving task, with real-time skeleton overlay and object tracking:
|
|
@@ -105,8 +82,6 @@ Display is a snapshot from our motion capture studio showing a subject performin
|
|
| 105 |
|
| 106 |
Watch a short demonstration of the motion capture data in action:
|
| 107 |
|
| 108 |
-
## π₯ Preview Video
|
| 109 |
-
|
| 110 |
<video width="100%" height="auto" controls preload="none">
|
| 111 |
<source src="assets/videos/jibengongjia.mp4" type="video/mp4">
|
| 112 |
Your browser does not support the video tag.
|
|
|
|
| 61 |
|
| 62 |
---
|
| 63 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 64 |
## π Data Format Specifications
|
| 65 |
|
| 66 |
| Component | Format | Details |
|
|
|
|
| 70 |
| Object 6D pose | `.csv` | Position (m) + quaternion per object, 120 Hz |
|
| 71 |
| Multi-view video | `.mp4` | 2β4 synchronized cameras, co-registered timeline |
|
| 72 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 73 |
## πΈ Sample Visualization
|
| 74 |
|
| 75 |
Display is a snapshot from our motion capture studio showing a subject performing a box-moving task, with real-time skeleton overlay and object tracking:
|
|
|
|
| 82 |
|
| 83 |
Watch a short demonstration of the motion capture data in action:
|
| 84 |
|
|
|
|
|
|
|
| 85 |
<video width="100%" height="auto" controls preload="none">
|
| 86 |
<source src="assets/videos/jibengongjia.mp4" type="video/mp4">
|
| 87 |
Your browser does not support the video tag.
|