S-JEPA: Joint Embedding Predictive Architecture for Self-Supervised Skeletal Action Recognition

A self-supervised method for skeletal action recognition based on the Joint Embedding Predictive Architecture (JEPA).

Overview

S-JEPA (Joint Embedding Predictive Architecture for Self-Supervised Skeletal Action Recognition) is a research project focused on learning strong skeletal representations without labels. The method instantiates JEPA for skeletal data to predict informative latent targets and improve downstream action recognition.

Authors

• Mohamed Abdelfattah
• Alexandre Alahi

Venue

ECCV 2024

One-Sentence Summary

S-JEPA is an instantiation of the Joint Embedding Predictive Architecture (JEPA) for self-supervised skeletal action recognition.

Links

• Project Page
• Paper

BibTeX

@inproceedings{abdelfattah2024sjepa,
  author={Abdelfattah, Mohamed and Alahi, Alexandre},
  booktitle={European Conference on Computer Vision (ECCV)},
  title={S-JEPA: Joint Embedding Predictive Architecture for Self-Supervised Skeletal Action Recognition},
  year={2024},
  organization={Springer}
}

Project Status: ✅ Published at ECCV 2024
Project Page: sjepa.github.io
Paper: S-JEPA