diema-challenge: body-motion emotion recognition (code only, no weights)

This page describes the model from the paper Orthogonal Ensembles and Tested Explanations for Performer-Independent Body-Motion Emotion Recognition (Naoto Nishida and Yoshio Ishiguro, arXiv:2609.02510). The code is at https://github.com/nawta/diema-challenge.

This repository contains no trained weights. The models were trained on the DIEM-A dataset, whose User License Agreement prohibits redistributing the data or any derivative of it. To reproduce the models, request access to the data and train with the code in the GitHub repository.

What the model does

It classifies a full-body motion-capture clip (a 25-joint skeleton, 64 frames) into one of 12 emotions. Each clip goes to 11 models from four families (graph convolution networks, attention models, hybrid/MLP models, and models with external pretraining), and the submitted system averages their softmax probabilities with equal weights.

Result

On the hidden test set of the DIEM-A Challenge at MMAC @ ACII 2026 (18 performers not seen in training, 1,944 clips), the system scored 37.23% Macro-F1, the unweighted mean of the 12 per-class F1 scores (chance is 8.3%). It received the challenge's Best Performance Award.

Try the code without the data

The GitHub repository includes tools/smoke_demo.py, which builds the models with random weights and runs them on random input. See "Try it without the data" in its README.

Citation

@misc{nishida2026orthogonalensemblestestedexplanations,
      title={Orthogonal Ensembles and Tested Explanations for Performer-Independent Body-Motion Emotion Recognition},
      author={Naoto Nishida and Yoshio Ishiguro},
      year={2026},
      eprint={2609.02510},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2609.02510},
}