Multi-task Self-Supervised Visual Learning

Carl Doersch, Andrew Zisserman

We investigate methods for combining multiple self-supervised tasks--i.e., supervised tasks where data can be collected without manual labeling--in order to train a single visual r…