Download page is here.
In order to facilitate the second challenge, we provide three datasets as follows.
Training set part one: this dataset contains the training images for the base set. This dataset is provided to build models with general facial perceptions. For example, to learn face representation, calculate average face, etc. In this dataset, there are 20K celebrities, each with 50-100 images. We selected these 20K celebrities from our 100K celebrity list and cleaned the dataset by running algorithms to remove outliers. To verify the data accuracy after cleaning, we randomly selected 2000 images from the base dataset and obtained the estimate of the labeling accuracy as about 99.5%.
Training set part two: this dataset contains the training images for the novel set. This dataset contains 1000 celebrities, who are NOT included in the 20K celebrity list in the training set one. There are limited number (1, 2, 5) of training images for each celebrity in this dataset. Participants are required to train models using the images in this dataset to recognize the same 1000 celebrities in the test set.
Test set Our measurement set contains a mixture of test images from both the base set and the novel set. We provide about 5 test images per celebrity in the base set while 20 images per celebrity in the novel set. We have run a de-duplicate algorithm to ensure that no test image is included in either of the training sets.
Notes about External Data Usage
In the low-shot learning challenge, you can use external data (especially for the base model training), but not the images for the persons in the low-shot classes. We restrict these persons in the low-shot classes only have limited (1,2,5) images for training for the sake of low-shot learning and fair comparison.
In order to remove the images for the persons in the low-shot classes, we provide the Freebase MID at the third column of the low-shot classes data. The names of the corresponding MID is located on this page.
Moreover, we encourage people to use the 20K celebrities in the base set to train the base model for easier comparison. According to our experiment, 20K celebrity can generate comparable results to the state of art on the LFW verification task.