We ask you to please not look at or use the provided test sets in any way other than for offline evaluation.
We ask you to only use the provided data for developing your solution (unless otherwise explicitly stated).
Submissions that rely on human intervention are allowed. The intervention strategy must be clearly explained such that the results are as reproducible and extensible as possible.
Algorithmic submissions may not rely on external intervention (e.g. humans, extra data). The results should be reproducible and extensible to other datasets.
Your developed solution should be practical and reasonably efficient given the scope of the challenge (e.g., your algorithm shouldn’t perform an exhaustive search).
Rules regarding participation:
Participants can only belong and participate in one team
Individuals are considered a team
Teams must be defined before the end of the challenge
Each team must have a leader who is responsible for submissions to the online evaluation platform
Participants should not access or inspect submissions or selection code from other participating teams until after the challenge concludes
Each training set that is part of the final submission will be limited to 1,000 data points. Training sets with more than 1,000 ( imageID, label) pairs will be rejected
The provided candidate pool has no labels, and as such, part of the challenge involves using the information contained in the embeddings as effectively as possible.
The provided candidate pool is a custom subset of the training set for the Open Images dataset. You may refer to metadata from the Open Images dataset [link]
If needed, you can leverage the human-verified and/or machine generated labels available in the metadata from the Open Images dataset [link]. However, we encourage creative solutions that minimize the amount of labels used.