Learning Privacy-Preserving Transferable Video Representations

Abstract

Pretraining on massive video datasets has become essential to achieve high action recognition performance on smaller downstream datasets. However, most large-scale video datasets are accompanied with issues related to privacy, ethics, and data protec-tion, often preventing them to be publicly shared with the community for reproducible research. Existing work has attempted to alleviate these problems by blurring faces, downsampling videos, or training on synthetic data. On the other hand, analysis on the transferability of privacy-preserving pretrained models to downstream tasks has been limited. In this work, we study this problem by first asking the question: can we pretrain models for human action recognition with data that does not include humans? To this end, we present, for the first time, a benchmark that leverages real-world videos with humans removed and synthetic data containing virtual humans to pretrain a model. We then evaluate the transferability of the representation learned on this data to a diverse set of downstream action recognition datasets. Furthermore, we propose a novel pre-training strategy, called Privacy-Preserving MAE-Align, to effectively combine synthetic data and human-removed real data. Compared to previous baselines, our approach reduces, by a large margin, the performance gap between human and no-human action recognition representations on downstream tasks. Our benchmark, code, and models will be made publicly available.