About

This grand challenge is focused on Heterogeneous Face Recognition (HFR), specifically on polarimetric thermal-to-visible matching. The motivation behind this challenge is the development of a nighttime face recognition capability for homeland security and defense. The challenge organizers will provide a polarimetric thermal and visible face database for algorithm development. Participants will be asked to submit heterogeneous face recognition algorithms that take a pair of images (an aligned polarimetric thermal face image and an aligned visible face image) as input and provide a similarity score as output. Algorithms will be ranked by their face verification performance using ROC curves.

Dataset

Upon signing and submitting a Database Release Agreement, participants will receive a multi-modal face database consisting of polarimetric thermal and visible imagery for algorithm development. Requests for the database can be made directly to the challenge organizers. This database consists of images from 60 subjects. Each polarimetric thermal face sample is represented by 4 images: the Stokes images S0, S1, S2, and the degree of linear polarization (DoLP). Additional details of this database can be found in the following reference:

S. Hu, N. Short, B.S. Riggan et al., “A polarimetric thermal database for face recognition research”, IEEE Conf. on Computer Vision and Pattern Recognition Biometrics Workshop, pp. 187-194, 2016.

Training Procedure: The organizers suggest that the participants develop/train their algorithms using two different protocols.

• Protocol 1: Train using imagery from 30 subjects, and test using imagery from the remaining 30 subjects. Use 5 random splits, to assess the average face verification performance as well as the standard deviation. This protocol will give the participant a sense of how well the algorithm is performing during development. Please submit the average receiver operating characteristic curve across the 5 random splits to the organizers for informational purposes – this will not be used in the evaluation/judging process.
• Protocol 2: Train using imagery from all 60 subjects. If the technique utilizes random initializations (e.g., such as in neural networks), please obtain five independent models corresponding to five different random initializations.

Challenge Evaluation

The organizers will use a sequestered dataset, similar to the provided training dataset, presenting pairs of visible and polarimetric thermal face samples representing both true match pairs and false matches pairs. If multiple models are submitted (if applicable, corresponding to random initializations), 5 match scores will be obtained for each test pair using the 5 submitted models/executable files, and averaged. Receiver operating characteristic curves representing face verification performance will be generated from the similarity scores. Participant performance will be judged/ranked based on area under the curve, equal error rate, and true match rate at false match rate of 0.01.

Submission

Each participant will be asked to deliver software in the form of an executable file that can be run either in a 64-bit Windows 10 architecture, or 64-bit Ubuntu architecture. Alternatively, the organizers will accept Matlab submissions, either m-file, p-code, or Matlab compiled .exe with MCR as part of the distributable package. Any executable submission should be fully encapsulated, with no external dependences. Each executable (or main Matlab function) should accept 6 inputs, and print the output to the terminal:

• Input 1: Full file path to a visible face image
• Inputs 2-5: Full file paths to a S0 image, a S1 image, a S2 image, and a DoLP image corresponding to a single polarimetric thermal face sample. Note that the participant is not required to utilize all polarimetric images (whichever combination that can be used to achieve maximal performance is acceptable), but the executable must accept inputs 2-5 as arguments.
• Input 6: Model number (1-5 corresponding to the random initialization, if applicable). If the approach does not involve random initialization, please inform the organizers – a value of 0 will be used for Input 6.
• Output: Floating point similarity score, corresponding to the degree of similarity between the visible and polarimetric thermal sample pair. Please print the output to the terminal. The submission should not print anything else to terminal.

The participant will also be requested to provide a brief write-up describing the submitted algorithm.

Every executable should be able to run from a terminal using the following format:

$ <exec_file> <Input 1> <Input 2> <Input 3> <Input 4> <Input 5> <Input 6>