For example, different self-supervision algorithms, hyper-parameters, unlabeled data.
Hello, thanks for reaching out. I am one of the challenge organizers. Can you clarify what do you mean by “different hyper-params, unlabeled data”?
Normally, for each track, you are free to use any self-supervision approach, self-supervised pre-training hyper parameters, unlabeled data as long as it’s used in self-supervised manner (see, Terms and Conditions here https://evalai.cloudcv.org/web/challenges/challenge-page/405/evaluation).
There are two choices:
- Train a self-supervision model specifically for each track.
- Train a single self-supervision model that is able to generalize to tasks in all tracks.
Obviously the latter is more challenging. I want to know if the former is allowed in this challenge.
This is a great question. We allow single model for each track. So option 1) is NOT allowed. We will update our challenge website as well to reflect this more explicitly.
In this way, the participants cannot participate in only one track, they must attend all the tracks. Is that right?
Participating in tracks is optional. If one wants to submit the results for only one track, they can do that. If they want to submit for multiple, that’s totally fine and in that case, use the same self-supervised model.