The recognition of occluded faces is a major challenge for face recognition and has received attention in an early stages of face recognition development. However, the problem has received relatively less representation in face recognition research and relative benchmark in comparison to issues related to cross-pose, cross-age, and cross-quality recognition.

The Advanced Occluded Face Recognition (IJCB-OCFR) competition aims to revive the topic of occluded face recognition and restart research efforts in this field. The main goal is to provide a challenging platform to explore different concepts that can be used to overcome the occluded face challenge, giving future direction and tools for researchers in the field. All participants who achieve competitive results will be invited as co-authors for a summary paper of OCFR, to be submitted to IJCB 2022.

Please read carefully all the following guidelines before registering and/or participating:

Task info:

1. The main competition will be based on verification between non-occluded references and occluded probes.

What to submit:

1. Participants will need to provide executables of their algorithms. The submitted algorithm should be a Win32 or Linux console application, that will need to run in the evaluation environment;
2. A template script will be provided. It will include code for reading the data and writing results, and each team will have to replace the processing method with their solution;
3. Participants will be prompted to submit answers to a set of questions regarding their submitted algorithms along with the submission;
4. Each team can submit up to 2 different solutions.

Evaluation environment:

1. Evaluation is going to be conducted on a NVIDIA Tesla V100 32 Gb GPU with the following CUDA version: V11.0.221;
2. A file with the complete systems' requirements can be found here;
3. Face bounding boxes for the evaluation set will be provided.

Evaluation and ranking criteria:

1. The baseline performance evaluation will be based on the open-source implementation of the ArcFace model. The considered model architecture is LResNet100E-IR trained on ms1m-refine-v2 database with ArcFace loss function. The pre-trained model is available on the official ArcFace Github repository;
2. The algorithms evaluation will be based on the verification performance. The verification evaluation will be based on the verification performance of occluded vs. non-occluded verification pairs, as this is the common scenario, where the reference is non-occluded, while the probe might be occluded;
3. The verification performance will be evaluated and reported as the false non-match rate (FNMR) at different operation points FMR100, FMR1000, and ZeroFMR, which are the lowest FNMR for a false match rate (FMR) <1.0%, <0.1%, and <0%, respectively;
4. The ranking of the submitted algorithms will be based on FMR100. If multiple submitted algorithms achieve the same FMR100 score, then FMR1000 will be considered, then the ranking will move to the separability between genuine and impostor comparisons measured by the Fisher Discriminant Ratio (FDR);
5. All submitted algorithms achieving FMR100 higher than the baseline verification performance (ArcFace) will be considered as a competitive solution.

Additional information:

1. The reference (enrol) images are generally not occluded, however some might have some minor occlusions (e.g. glasses);
2. The probe (test) images can have various levels of occlusions, starting from almost no occlusion to strongly occluded;
3. The occlusions can be caused by any undisclosed factors and they can be located in any part of the face;
4. The final competition paper will likely also report occluded vs. occluded results.

Cap on inference time/complexity:

1. Up to 3x baseline inference time on the testing environment.

Any issue not covered here should be reported to the organisers, who will be resposible for deciding on the most appropriate solutions.