lei-qi-233/MicroG-4M-models

MicroG-4M: Human Action Recognition in Microgravity

This repository contains fine-tuned weights for the Human Action Recognition (HAR) task, as presented in the paper Go Beyond Earth: Understanding Human Actions and Scenes in Microgravity Environments.

MicroG-4M is the first benchmark for spatio-temporal and semantic understanding of human activities in microgravity. It covers 4,759 clips across 50 action categories from real-world space missions and simulations, addressing the gap in domain-robust video understanding for safety-critical space applications.

Resources

• Paper: Go Beyond Earth: Understanding Human Actions and Scenes in Microgravity Environments
• GitHub: HAR-in-Space
• Dataset: MicroG-4M on Hugging Face

---

Performance comparison of models fine-tuned on MicroG-4M for HAR

Arch	TC	Backbone	#Params (M)	mAP (%)	F1-score (%)	Recall (%)	AUROC (%)
C2D	8×8	R50	23.61	29.51	8.09	6.58	83.49
C2D NLN	8×8	R50	30.97	44.64	28.30	24.86	89.40
I3D	8×8	R50	27.33	46.41	26.37	22.25	88.79
I3D NLN	8×8	R50	34.68	47.12	28.07	24.65	88.52
Slow	8×8	R50	31.74	45.19	26.13	22.77	88.49
Slow	4×16	R50	31.74	46.37	28.72	25.38	88.30
SlowFast	8×8	R50	33.76	43.02	22.63	18.98	88.51
SlowFast	4×16	R50	33.76	42.09	23.69	20.18	87.54
MViTv1	16×4	B-CONV	36.34	12.86	5.54	4.66	74.63
MViTv2	16×4	S	34.27	15.14	8.16	7.17	78.61
X3D	13×6	S	2.02	14.07	5.77	4.52	78.23
X3D	16×5	L	4.37	18.70	9.15	7.47	78.27

Note:

• All models have been pretrained on the Kinetics400 dataset and continually trained on MicroG-4M.
• TC denotes the temporal configuration (frame length × sampling rate).
• #Params indicates the number of parameters (in millions, M).

Contents of this repository:

• models folder contains all fine-tuned weights of MicroG-4M
• MicroG-4M_results folder contains all raw data generated by fine-tuning

Citation

If you find this work useful, please cite the following paper:

@misc{wen2025earthunderstandinghumanactions,
      title={Go Beyond Earth: Understanding Human Actions and Scenes in Microgravity Environments}, 
      author={Di Wen and Lei Qi and Kunyu Peng and Kailun Yang and Fei Teng and Ao Luo and Jia Fu and Yufan Chen and Ruiping Liu and Yitian Shi and M. Saquib Sarfraz and Rainer Stiefelhagen},
      year={2025},
      eprint={2506.02845},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2506.02845}, 
}