Problem
Extracting tailoring-grade (~0.25-inch) real-world body dimensions from a phone video, without industrial hardware like lasers.
Approach
A "triple-lock" scaling strategy: IMU physics for metric motion, a known reference object (A4 paper) to anchor the coordinate system, and the SMPL-X statistical body model to infer measurements hidden by clothing or perspective.
What I built
- Camera intrinsic-matrix calibration and metric-depth integration (ZoeDepth / Depth Anything V2).
- Visual-inertial odometry syncing 100 Hz IMU data with 30 fps video to recover the camera’s 3D path.
- SMPL-X mesh fitting with cross-sectional slicing for girth measurement; ~4.7k LOC with tests and analysis.
Result
Still in progress. The optical calibration and capture flow work, and rigid-object reconstruction runs end to end; the body-measurement and regression-correction layers are still in development.