@jason_brumwell If the list of test images is known in advance and hand-labeling is allowed what prevents us from hand-labeling of test images and getting a perfect score (with model training on these labels)? It sounds like this approach is not prohibited by the rules, but it is not the solution you are looking for (as far as i can understand).
Thank you for the challenge, I have a few questions:
Is it ok to manually hand-label the examples from the provided dataset (with 2200 groups) if a trained model will generalize the data with unknown teams added (dataset with 22000 groups)?
What does the score/secondary score columns on the leaderboard means? According to the description, “score” is a quality on the dataset with 2200 groups, “secondary score” is a quality on the dataset with 22,000 groups, but the rules state: “The results achieved must not deviate by more than 5% when run against datasets with different team images.”. Does it mean that must be gap no more than 5% between score/secondary score columns on the leaderboard?