MocapLab

Two R&D internships in machine learning applied to motion capture — detecting eye and facial tracking errors, then automatically deduplicating motions for a sign-language dictionary.

The studio

MocapLab is a Paris-based studio entirely dedicated to motion capture, founded and led by Rémi Brun — the founding team started doing mocap as early as 1993, on the pilot for20,000 Leagues Under the Sea, one of the very first feature films to use the technology. The studio covers four capture families (body, face, eye, fingers) and claims over 300 completed projects across film, TV, video games and live entertainment, including cinematics for Syberia and capture work for Ubisoft's Riders Republic.

Beyond service work, MocapLab runs an internal research effort, notably aroundSign3D, a 3D-avatar app for French Sign Language funded by France 2030 — a project that draws directly on the studio's finger-capture expertise.

mocaplab.com →

R&D Intern

03/2026 – 07/2026

Context

Eye-tracking motion capture data (TRC files, high-frequency facial markers) contains tracking errors — occlusions, marker slips, rigidity loss — that today have to be spotted by eye by an operator before any use in production.

Approach

Analyzed which errors cost production the most, then trained several tracking and automatic-labeling models for the eye markers in Python — supervised classifiers, regression, and a CNN compared on the same feature set to pick the right complexity for each error type, with a decision threshold tuned to favor precision: a wrongly flagged marker costs more downstream than one left undetected.

Result

Model integrated into the team's production pipeline, as an aid to spotting errors rather than a replacement for human review.

Intern

09/2025 – 10/2025

Context

A sign-language dictionary built from motion capture accumulates near-identical variants of the same sign, recorded at different times — with no prior labeling of duplicates.

Approach

Designed a Python ML model for motion classification and automatic deduplication: group captures that correspond to the same sign despite natural execution variation, keeping only one representative entry per sign.

Result

Results and methodology presented to the research team at the end of the internship.

← back to personal projects