Recognizing One Million Celebrities
There are several datasets for measurement purpose. To clarify, we summarize their descriptions in the following table.
|Dataset||Purpose||Face Aligned or Not||Random/Hard|
|Face-Aligned-Dev-1||Development (label provided)||Face aligned||Hard selection|
|Face-Aligned-Dev-2||Development (label provided)||Face aligned||Random selection|
|Face-Cropped-Dev-1||Development (label provided)||No alignment||Hard selection|
|Face-Cropped-Dev-2||Development (label provided)||No alignment||Random selection|
|Face-Aligned-Test-(part 1)||Test (no label provided)||Face aligned||Hard selection|
|Face-Aligned-Test-(part 2)||Test (no label provided)||Face aligned||Random selection|
|Face-Cropped-Test-(part 1)||Test (no label provided)||No alignment||Hard selection|
|Face-Cropped-Test-(part 2)||Test (no label provided)||No alignment||Random selection|
For the dataset with the “development” purpose, we release the ground truth label. Participants can use this dataset to develop and tune their algorithms. The dataset with the “development” purpose could be downloaded here.
For the datasets with the “test” purpose, we do NOT include the ground truth label. 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).
Code for Evaluation
python PrecisionCoverageEst.py -config Curve_Development.config -pthresh 0.95 This sample results file has the following format
- Column1: Freebase MID (ground truth)
- Column2: EntityNameString
- Column3: ImageURL
- Column4: FaceID
- Column5: MID estimated
- Column6: Confidence score
Participants can use this code to test the performance of their algorithms on the development set.
To obtain results with the test set, participants need to upload their results after they sign in here. 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
If participants use their own alignment algorithm, we recommend to use the dataset with the “no alignment” option and run their own alignment algorithm. If participants use our aligned face images for training, we recommend to use the dataset with the “face aligned” option. The “hard selection” and “random selection” is described here.