First of all I would like to thank mmaction2 for bringing all SOTA activity recognition algorithms together. Because of this library I was able to propose my work on a custom dataset.

In the suggestions, during my proposal, I was asked to explore Spatio-temporal proposals. I see that there are two methods available for this in mmaction2.

The question I have is regarding the "region proposals". I need to implement it for my custom dataset so I want to make sure I got this right,

1. The proposal boxes + time stamps are from FAIR's Long-term Feature Banks
2. slowfast/slowonly networks (trained as recognizer on the AVA dataset) are used to predict the activity in those proposals.

Question: Is the slowfast/slowonly networks retrained on the proposals from FAIR's long-term feature bank?

Please let me know if I don't make any sense. I will try to formulate my question more carefully.