Welcome to CyberuSentry! CyberuSentry is a mixture of 4 words that describe the nature of this project.

• Cyber - Involving, using, or relating to computers
• Cerberus - A three-headed guardian dog from the Greek mythology
• Sentry - A soldier who guards a place, usually by standing at its entrance
• Entry - The act of entering a place

In this repository you can find the implementation for the CyberuSentry model. CyberuSentry is an ensemble of siamese neural networks developed for the purpose of participating in the UCCS Watchlist Challenge. You can find the summary paper on this link, and relevant poster here

The model unfortunately placed last in the challenge. However, such a benchmark conducted by a third-party organization provides valuable insight into the current state of the model, which is one of the main reasons why I decided to sign up for the challenge. Overall, I am satisfied with the result.

You can find the evaluation results in the next section. For detailed description of my approach, see the PDF report

The task of this challenge, organized by the University of Zurich, was to develop a model capable of solving open-set face identification problem. There was also a part dedicated to face detection, though this part was not considered within this model and pre-trained face detectors were used to facilitate work in this repository.

The competition started on March 1, 2024. I signed up after I found the competition was ongoing and my application was approved on March 22, 2024. The deadline for submitting all the necessary files was June 1, 2024.

Since only four teams participated in the challenge, the results are a little scarce. However, as the other participating teams included, among others, eg. Federal Police of Brazil, I assume the participant field was very competetive. In retrospective, I am satisfied with the overall results, as I was developing the project completely alone, from scratch, in my free time and using my limited computational resources. Just the fact that I managed to develop a working facial recognition system in two months is considered a success in my eyes. I am confident that I could improve the results, had we been given some more time.

You can find the evaluation curves below. Models from my submissions are those beginning with the prefix TR-. For more details, please refer to the summary paper

Since the UCCS Dataset is not public at the moment, the trained models will not be released as every single one of them was trained using proprietary data. Should the dataset become public in the future, I will reconsider publishing the models again.

• Images were generated using Gencraft website. All images used were created under a paid subscription.
• Cambridge Dictionary
• Sample video from Pexels
• UCCS Watchlist Challenge
• Summary paper on Arxiv
• Poster presented at IJCB 2024