[2Yin-B] Poster 4

[2Yin-B-31] Generalized Performance Representation Learning in Sports Motion via Bioinformatics

〇Tomohiro Suzuki¹, Taichi Hibi¹, Teppei Shimamura², Keisuke Fujii^1,3, Norimasa Kobori^1,4 (1. Nagoya University, 2. Institute of Science Tokyo, 3. RIKEN, 4. Mercari, Inc.)

Keywords:

Motion Analysis,Representation Learning,Bioinformatics,Deep Generative Models

The evaluation of sports motions has often relied on features defined according to the sport, motion type, and individual physical characteristics. This made it difficult to compare skill levels across different motions or individuals, or to extract common performance factors. This study applies scANVI, a single-cell analysis method from bioinformatics, to acquire a general motion performance representation independent of motion type. Specifically, it trains a model by treating the time-series motion data obtained from 3D joint coordinates as "cells," the motion type and subject ID as "batch effects," and the performance evaluation of the motion data as "cell types." The proposed method employs a deep generative model incorporating ordered regression to generate latent representations that preserve performance order while eliminating action category information for track and field motions. Experimental results show that within the acquired latent space, action categories become mixed, yet performance levels are clearly separated and aligned. This suggests the proposed method enables the extraction of an abstract feature, "motion proficiency," independent of action type.