Body Dimension Estimation from Clothed 3D Scans
Research at the University of Michigan Transportation Research Institute under Dr. Byoungkeon Daniel Park. Built a standalone tool that recovers body measurements from scans of clothed subjects, so studies no longer need subjects scanned in fitted clothing.
The recovered research artifacts document a nine-view pipeline. A 3D scan was rendered at 45° intervals from 0° through 360°. OpenPose produced 25 joints per view, represented as x, y and confidence—75 values per image and 675 across the nine renders. Accepted 2D detections were ray-picked back onto the scan and transformed into the original 3D coordinate system before measurement.
The interactive viewer is a privacy-safe reconstruction of that process. Its point cloud is generated from equations in the browser, and its dimensions are illustrative. No participant scan, photograph, research output, or confidential code is included.
The artifacts also explain two useful debugging lessons: an early version captured eight images while downstream code expected nine, and the C# and Python implementations initially disagreed about transformation order. Making the camera count and coordinate frames visible was as important as the pose model itself.