What is the current state of the art in video transformers (mainly for tasks like classification) and what are the Top 5 papers from the last 2 years?

Question

Is there a general consensus in the community regarding

• the most effective video transformer architecture
• which modalities to use, how to represent them, and the best methods for fusing them
• the recommended training strategies

for tasks like video classification

I don't really know where to start. Of course, I’ve reviewed papers from major conferences, but there are so many and they are partly very specific, so it is hard for me, as someone rather new to the field, to evaluate which works are most relevant.

I am happy about every resource you recommend for learning about this.

Thanks!