The training database consists of a set of folders, where each folder contains pictures of a single person. You are free to sample pairs of same / different persons and learn to distinguish between the two cases. Also any other training data is allowed (but not mandatory) for this purpose (e.g. LFW or YTF datasets). The training database also contains the following readily calculated data for each image:
\r\n<image_name>_landmarks_big.csv - 68 landmark coordinates extracted from the image.
\r\nThe test database consists of images similar to those in the training database. However, the identities are not given. Instead, a list of image pairs (the third file) is provided, and competitors are required to determine whether each pair is a match (1) or non-match (0).
\r\nNote: The data was re-uploaded on October 28th, because the deep features were incorrectly computed. Please reload train.zip and test.zip.
", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Description", "content": "Face verification is a widely studied research topic, where the goal is to automatically determine whether two images represent the same person. Until recently, the dominating techniques used facial landmarks as the features for identity verification. Typical landmarks include the coordinates of eyes, nose, corner of mouth, etc. More recently deep learning based techniques have changed the domain by reaching human level accuracy (97 % and above) in the verification process [1,2].
\r\nThe goal of this competition is to develop algorithms for accurate face verification. The goal is to learn to predict whether a pair of images represent the same person (target output = 1) or a different person (target output = 0). In order to help you get started, we've prepared three sets of possible features: two sets of landmark locations and one set of deep features calculated using the AlexNet as described here.
\r\nPrizes: The best teams from TUT and TUGraz will be donated a T-shirt for each team member. Moreover, the winning team will receive shirts as well, if not from the related courses of the organizing universities. We'll try our best to get hold of XKCD shirts, although they are in high demand. If not possible, we'll find another nerd T-shirt instead.
\r\nOrganizers: Gernot Riegler (Graz University of Technology, Austria) and Heikki Huttunen (Tampere University of Technology, Finland).
\r\n[1] Taigman, et al. \"Deepface: Closing the gap to human-level performance in face verification,\" IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014.
\r\n[2] Schroff, et al., \"FaceNet: A Unified Embedding for Face Recognition and Clustering,\" IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015.
", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "rules", "content": "\r\n\t\r\nYou cannot sign up to Kaggle from multiple accounts and therefore you cannot submit from multiple accounts.
\r\n\r\n Privately sharing code or data outside of teams is not permitted.\r\n It's okay to share code if made available to all participants on the forums.\r\n
\r\nTeam mergers are not allowed in this competition.
\r\n\r\nThe maximum size of a team is 3 participants.
\r\nYou may submit a maximum of 4 entries per day.
\r\nYou may select up to 2 final submissions for judging.
\r\nStart Date: 10/14/2015 6:25 PM UTC
\r\nMerger Deadline: None
\r\nEntry Deadline: None
\r\nEnd Date: 12/11/2015 11:59 PM UTC
\r\n\t\r\nUse of additional training data is allowed (e.g. different datasets to train your classifier).
\r\nManual annotation of identities in the test data is not permitted. The winners are required to submit their code for inspection.
\r\nTUT & TUGraz students: Include an identifier into your team name (e.g. at TUT; your student number such as TeamASDF (143712, 143713, 143714)).
", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Submissions And Scoring", "content": "The participants will be ranked based on the accuracy of their predictions. The accuracy is measured by Weighted AUC, which assesses the accuracy based on the area under the Receiver Operating Characteristics (ROC) curve. This measure is suitable for cases where the classes are severely imbalanced: I our competition only about 0.3% of pairs in the test data are a match. With a more straightforward measure--such as the prediction error--the all-zeros prediction with 99.7% accuracy would be difficult to beat.
\r\nThe AUC ranks the submissions based on the order of predicted likelihoods. The prediction should consist of not only the predicted category (match or no-match), but also its likelihood (score between 0 and 1). It is allowed to submit only zeros and ones, but the score will be less than with well justified likelihoods represented as real numbers.
\r\nThe submission file should be a text file with the following contents:
\r\nId,Prediction\r\n
0,0.523359
1,0.211905
2,0.408814
...
901151,0.803045
901152,0.677751
The first column of the csv file is the index of the pair of images (see file pairs.csv). The second column is the likelihood that the images are a pair: higher values imply higher confidence.
", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false } ] }