shun31y/ELMA-sample

ELMA Episodic Motion Dataset Sample

This repository provides a small sample subset of ELMA.

It consists of 30 randomly selected episodes from the ELMA dataset. For each episode, we provide the original GVHMR output in .pt format, the converted SMPL-X / HumanML3D-style .npz file for visualization with the Blender add-on, and a .gif visualization that combines the motion with its corresponding text description. In the .gif, the caption of each 5-second segment is shown in red when the segment refers back to an earlier one (event anaphora).

The dataset is organized into the following directories:

.
├── hml3d_gif/
├── raw/
└── smplx_npz/

• Full dataset: https://huggingface.co/datasets/shun31y/ELMA
• Code / baselines: https://github.com/shun31y/elma-benchmark
• License: CC BY-NC 4.0

Citation

@inproceedings{kato2026elma,
  title     = {{ELMA}: Benchmarking Anaphoric Compositions for Long-Term Text-to-Motion Generation},
  author    = {Kato, Shun and Yang, Dixin and Amaya, Kotaro and Wada, Yuiga and Fukuda, Kazumi and
               Shibuya, Takashi and Mitsufuji, Yuki and Isogawa, Mariko},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
  year      = {2026},
  eprint    = {2610.13549},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}