Question Regarding the Training Strategy

[vedantdalimkar]

Hi @LiheYoung, I was going through one of your previous works - ST++, where you emphasise on a image-selection approach over pixel-selection approach for SSL. However, in Unimatch you have gone with the pixel-selection approach.

Also, Unimatch uses the online training paradigm instead of the offline retraining strategy proposed in ST++.

It would be very helpful if you could give me some insight on the rationale behind these 2 design choices in Unimatch.

[LiheYoung]

In UniMatch, we just follow the philosophy of FixMatch: simple, one-stage training. We did not explore some alternative more sophisticated choices such as image-level selection.