We prepare both the development set and test set as follows.
|Dataset||Purpose||Face Aligned or Not|
|Development Set||Development (label provided)||Face aligned|
|Test Set||Test (no label provided)||Face aligned|
For the development set, participants can download here. The development set is a tsv file, in which each line has the following format (delimited by tab).
- Column1: Base set or novel set
- Column2: Image ID
- Column3: FaceData_Base64Encoded
- Column4: Freebase MID
- Column5: ImageURL
For the test set, participants need to register here to receive the download link for these datasets. We will send emails with download link one week before the final results announcement (tentatively July 14th) -> evaluation due is changed to July 17th. Please provide the experimental results in a tsv file, in which each line has the following format (delimited by tab).
- Column1: LineNumberIndex (please use the line number we offered in the test file)
- Column2: MID estimated
- Column3: Confidence score
Note that since we focus on low-shot learning in this challenge, we only provide the aligned face option for easier comparison.
We will rank all the participants according to their performance on the novel set, while monitor the performance on the base set. More specifically, first, participants need to achieve at least 98% top-1 accuracy on the base set to make their results valid. Second, the teams will be ranked according to the coverage at precision = 99% on the novel set. (Note: evaluate with the one example per novel class option.)
Please note this year, if the participants achieve 99% top-1 accuracy on the base set, their score will be ranked separately. Next year, we will improve the minimum requirement of top-1 accuracy on the base set set to be 99%. Since this year is the first time we hold this low-shot learning challenge, we lower the minimum requirement to 98% to encourage more participants.