Dear organizers, I have several questions about the challenge. - Is it possible to apply test time adaptation on mini-batch or the entire test data? - If so, how does the input directory hierarchy look like? - Does the "domain unaware" task limit only the supervised information (labels) and permit unsupervised learning in the target domain? - Does the "domain unaware" task prohibit the use of public data in the target domain? * Is it possible for me to assess my model's performance on test data if I submit a docker container for each task? (I'm uncertain because a public leaderboard does not appear to be available.) * How much GPU memory and RAM are provided for the docker container? * What is the maximum number of times that I can submit the docker containers? * When is the precise deadline for the challenge? Does "end of September" mean 30th September in Anywhere on Earth time? Thank you. :)

Created by Jihun Yoon nicezebra
Sorry I only saw the thread just now Regarding time adaptation: Not exactly sure what you mean, but the model should look at each image independently Domain unaware: You are allowed to use the data from all domains for training, we just won't tell you the domain of a given image, so also any public data is fine Results on test data will only be provided on the challenge day. Good question, in previous challenges we never had any issues here, please make sure your model can run on a v100 you can submit as many as you want, we will score the final submission (unless you submit completely distinct methods). Good point, I will update this. Please submit until the 1st of October anywhere on earth

Questions about model training, evaluation, and submission page is loading…