AthletePose3D Validation Methodology
This document outlines the theoretical and technical methodology for validating Kinemotion using the AthletePose3D (AP3D) dataset, specifically focusing on Coordinate Normalization and 3D Joint Angle Calculation in monocular video analysis.
1. Overview
Validating a monocular (single-camera) pose estimation system against a 3D gold-standard dataset like AthletePose3D requires more than simple point comparison. To achieve research-grade results, we must account for perspective, distance, and the inherent limitations of 2D image data.
2. Coordinate Normalization (Pixels to MM)
Why it matters for Monocular Video
In monocular video, the same movement (e.g., a 50cm jump) results in different pixel displacements depending on:
- Camera Distance: An athlete further away produces fewer pixels per cm.
- Resolution: 1080p vs. 4K changes the pixel count for the same physical movement.
Methodology: Procrustes Alignment
To compare MediaPipe landmarks (normalized [0, 1] or pixel coordinates) with AP3D ground truth (world coordinates in mm), we use Procrustes Analysis (specifically the Kabsch Algorithm).
- Process:
- Translation: Center both point clouds (MediaPipe and Ground Truth) at the same origin (e.g., the midpoint of the hips).
- Scaling: Calculate a global scaling factor based on stable anatomical anchors (e.g., torso length or hip width) averaged over the entire sequence.
- Rotation: Apply a rotation matrix to align the orientations of the two models.
- Result: This yields a unit-invariant MPJPE (Mean Per Joint Position Error) that tells us how many physical millimeters the system is off by, regardless of camera placement.
Implementation Difficulty: Medium
- Math: Standard linear algebra.
- Challenge: Bone length jitter. MediaPipe bone lengths can “stretch” during fast movements; the normalization must use robust averaging to prevent the scale from shifting frame-by-frame.
3. 3D Joint Angle Calculation
Why it matters for Monocular Video
Traditional 2D analysis measures angles “flat” on the screen. This is subject to Perspective Error (Foreshortening):
- A 90° knee bend viewed from a 45° angle might appear as 70° or 110° in 2D.
- 3D Reconstruction: By using MediaPipe’s estimated Z-depth, we can reconstruct the true anatomical angle.
Methodology: Vector Analysis in 3D
We calculate the angle $\theta$ between two 3D vectors (e.g., Thigh and Shank) using the dot product:
\[\\cos(\\theta) = \\frac{\\mathbf{A} \\cdot \\mathbf{B}}{|\\mathbf{A}| |\\mathbf{B}|}\]
Where:
- $\mathbf{A}$ is the vector from Knee to Hip.
- $\mathbf{B}$ is the vector from Knee to Ankle.
Implementation Difficulty: Medium-Hard
- The Math: Straightforward.
- The Challenge: Depth Noise. MediaPipe’s Z-axis (depth) is its most unstable dimension. Small errors in depth (e.g., 2-3cm) can cause massive swings in the resulting 3D angle (15°+).
- Solution: Requires advanced signal processing (Savitzky-Golay filtering or Kalman filters) applied to the 3D vectors before angle calculation to ensure the metrics are stable enough for coaching feedback.
4. Strategic ROI
| Feature |
Strategic Value |
Marketing Claim |
| Normalization |
Proof of absolute accuracy. |
“Validated to within ±X mm of gold-standard research.” |
| 3D Angles |
Perspective invariance. |
“Accurate biomechanical angles from any camera view.” |
5. Summary
Implementing these methodologies moves Kinemotion from a qualitative analysis tool to a quantitative biomechanical instrument. While 2D analysis is sufficient for the Phase 1 MVP, these 3D refinements are the foundation for Phase 2 research-grade validation.