Unlocking Exocentric Video-Language Data for Egocentric Video Representation Learning

Zi-Yi Dou, Xitong Yang, Tushar Nagarajan, Huiyu Wang, Jing Huang, Nanyun Peng, Kris M. Kitani, FU-JEN CHU

Keywords: video-language pretraining, egocentric video

TL;DR: We present a framework that constructs video-language data from exocentric sources for egocentric video representation learning.

Abstract: We present EMBED (Egocentric Models Built with Exocentric Data), a framework designed to mine video-language data from exocentric sources for egocentric video representation learning. Large-scale exocentric data covers diverse activities with significant potential for egocentric learning, but inherent disparities between egocentric and exocentric data pose challenges in utilizing one view for the other seamlessly. In this study, we propose leveraging hand-object interactions and language narratives as cues to incorporate exocentric data into egocentric training. Specifically, we focus on identifying specific video clips that emphasize hand-object interactions and pairing them with action-focused language narrations. By applying our framework to exocentric datasets such as HowTo100M, we construct datasets thar are effective for egocentric video-language pretraining. Our extensive evaluations reveal that EMBED achieves state-of-the-art performance across various egocentric downstream tasks, including a 4.7\% absolute improvement in multi-instance retrieval on the Epic-Kitchens-100 benchmark and a 6.2\% improvement in classification on the EGTEA benchmark in zero-shot settings. Furthermore, EMBED enables egocentric video-language models to perform competitively in exocentric tasks. Finally, we showcase EMBED's application across various exocentric datasets, exhibiting strong generalization capabilities when applied to different exocentric datasets.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 12389