I made a 3D human mesh reconstruction & frame-sync web app for sports analysis
retained text
Technical stack info I can share about the 3D pipeline (works on regular handheld videos): **Bounding Box & Target Tracking:** MediaPipe JS library running in the browser for lightweight 2D pose tracking, used primarily for accurate automated cropping and keeping the subject centered for downstream processing. **3D Mesh Recovery:** A specialized 3D reconstruction model optimized for video sequences (operating on the cropped sequence), combined with a custom gravity and Inverse Kinematics (IK) pipeline to lock world orientation and adapt ground contact constraints. I use Meta’s MHR for the body model for an efficient body mesh model. **Temporal Alignment:** Fast approximate Dynamic Time Warping (DTW) to handle frame synchronization across movements with different tempos/speeds. **Rendering:** [Three.js](https://threejs.org/) for 3D rendering. [Activeframe](https://github.com/activetheory/activeframe) for fast video scrubbing. You can try it out for free at[ twinreplay.com](https://twinreplay.com/) with your own clips or check out the built-in interactive demos. Would love to hear your thoughts or feedback.
document fetched 2026-11-19 00:23 · canonical https://reddit.com/r/computervision/comments/1xzss3r/i_made_a_3d_human_mesh_reconstruction_framesync