Learning from Streaming Video with Orthogonal Gradients
| Properties | |
|---|---|
| authors | Tengda Han, Dilara Gokay, Joseph Heyward, Chuhan Zhang, Daniel Zoran, Viorica Pătrăucean, João Carreira, Dima Damen, Andrew Zisserman |
| year | 2025 |
| url | http://arxiv.org/abs/2504.01961 |
Abstract
We address the challenge of representation learning from a continuous stream of video as input, in a self-supervised manner. This differs from the standard approaches to video learning where videos are chopped and shuffled during training in order to create a non-redundant batch that satisfies the independently and identically distributed (IID) sample assumption expected by conventional training paradigms. When videos are only available as a continuous stream of input, the IID assumption is evidently broken, leading to poor performance. We demonstrate the drop in performance when moving from shuffled to sequential learning on three tasks: the one-video representation learning method DoRA, standard VideoMAE on multi-video datasets, and the task of future video prediction.