Training Data

The training set comprises the watchlist images and the probe images of 90 subjects. In the watchlist, 3 images of the subject's head are provided: 1) frontal image; 2) left-side image; 3) right-side image. In the probe images, 5 images of each subject are provided.

The identity of each image is provided in the image filename with following structure:

   ID_type.jpg

Example for the 3rd subject in the watchlist

   003_r.jpg - the right side watchlist image

   003_l.jpg - the left side watchlist image

   003_f.jpg - the frontal watchlist image

   003_01.jpg - the 1st probe image

   ...

   003_05.jpg - the 5th probe image

---

Meta-data

Considering that, in some cases, images contain multiple subjects, the face location of the interest subject is provided as a bounding box. These data was automatically inferred using a state-of-the-art head-landmark localization method [1] and corrected manually. The annotations are available in a text file where each line contains the information of an image according to the following structure:

   imageFilename X Y W H

X is x coordinate of the top left corner of the bounding box

Y is y coordinate of the top left corner of the bounding box

W is the width of the bounding box

H is the height of the bounding box

Example for the 3rd subject

   003_01.jpg 45 78 100 200

   003_02.jpg 450 306 490 524

---

Competition Task

Let W={w₁,w₂,...,w_N} be the set of subjects in the watchlist, G be the set of training images and P the set of images in the test set.

The participants are expected to provide an algorithm capable of performing the following task:

• Given a path with k probe images, the algorithm should output to a text file a matrix of k by N similarity scores, where the element at (i,j) position should contain the similarity score between the probe image P_i and the j^th subject in W.

Example of the application functioning, through command-line:

•    > mainApp 'path_to_probe_images' 'path_to_annotations_file'

The elements of the expected output text file should be separated by a white-space, and each line should contain the similarity scores of each probe image.

Fig. 3: ICB-RW fundamental task.

---

Evaluation

The algorithm performance will be determined by the area under curve (AUC) of the cumulative match score (CMC) curve. For this purpose, 5 probe images will be used, which are disjoint from the ones used by the participants.

For each probe image P_i , a rank-K list is constructed by selecting the K most similar watchlist subjects. The CMC curve relates the percentage of correct identification for all probe images with the size K of the rank-K list.

Fig. 4: The CMC curves used to rank participants.

---

Technical Notes

• The application executable can be written in any programming language (Matlab inclusively) and should run in one of the operating systems: ”Windows 7, Service Pack 1” or ”Ubuntu 14.04”.
• There will be no internet access during the evaluation. Thus, the application executable will need to be installed and executed without access to the internet.
• Each participant is allowed to submit one single algorithm and executable.
• The application should be submitted by email to icbrw@di.ubi.pt.