What markerless motion capture taught me about evaluation honesty
Metrics that win a slide can still lose the lab. Notes from a computer-vision thesis path.
The most dangerous number in a research paper is the one that looks best in the abstract.
For my TDK research at the University of Debrecen, I ran a systematic comparison of six AI-driven markerless motion capture systems — OpenPose, MediaPipe, FreeMoCap, Rokoko Vision, Move.ai, and Autodesk Flow Studio — against the Motion-X++ dataset, across basic, complex, and deliberately challenging movement categories.
The slide metric vs the lab metric
Every system had a category where it shone and a category where it fell apart — occlusions, fast lateral motion, unusual camera angles. Reporting only the flattering average would have made a better poster and a worse paper. The honest table, the one with the failures left in, is what turned the study into a guidance framework for practitioners choosing a system.
A benchmark is a promise you make to the next researcher.
What the committee rewarded
The work scored 46/50 at the TDK conference and was nominated for the National round. I am convinced the evaluation section — the part that admitted where each system breaks — is what carried it. Reviewers trust numbers that cost the author something.
This is sample copy for the article template. Replace with the final essay when you publish.
← All stories