Questions for Vdeo Representation Learning: Self-supervised Co-training for Video Representation Learning
---------------------------------------------------------------------------------------------------------

Please send your answers to David Hoffmann before 10:00 28.01.2021

1) Briefly explain the main steps of the proposed training algorithm and identify the novelty of the proposed approach. (2-3 sentences)

2) Despite the inferior accuracy that is empirically observed when optimizing the flow and RGB network simultaneously can you think of any other reason to prefer the alternating training procedure? (1-2 sentences)

3) Is this approach generally applicable and will lead to similar improvements for all possible (action recognition) classes and datasets? Where do you see limitations? (2 sentences)