MPS Hand Tracking Algorithm Performance Benchmark
This page presents the performance evaluation of the MPS Hand Tracking algorithm for Aria Gen1 glasses.
Benchmark Dataset
3D evaluation is conducted on the HOT3D Aria split, a publicly available dataset for hand-object interaction research.
- Dataset: HOT3D
- Total Duration: 400 minutes of Aria Gen1 recordings
- Number of Participants: 19
- Ground Truth Source: Motion Capture (MoCap)
Evaluation Metrics
We report the following metrics (lower is better for both):
| Metric | Description |
|---|---|
| LTR (Lose Track Ratio) | The ratio of frames where the hand is not tracked successfully. |
| MKPE (Mean Keypoint Error) | The mean 3D distance error over all hand keypoints and all frames, measured in millimeters. |
Results
| MPS HT Version | LTR (%) ↓ | MKPE (mm) ↓ |
|---|---|---|
| MPS HT 2.0.0 | 6.31 | 51.97 |
| MPS HT 3.1.1 | 3.94 | 22.80 |
About HOT3D Dataset
HOT3D is a large-scale dataset for egocentric hand and object tracking in 3D. The Aria split of HOT3D provides diverse hand-object interaction scenarios captured with Aria Gen1 glasses, making it an ideal benchmark for evaluating hand tracking algorithms in real-world conditions.
For more information about the HOT3D dataset, visit the Project Aria Datasets page.