Large group face recognition for attendance tracking

• Main Author: Yeap, Jia Hao
• Other Authors: Ong Chin Ann
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Attendance tracking; Facial recognition
• Online Access: https://hdl.handle.net/10356/175039
• Institution: Nanyang Technological University

Large group facial recognition for attendance taking has a very practical application use case in today’s academic institutions. With the multitude of modern state-of-the-art models being developed, this project will serve to evaluate the feasibility of large group facial recognition for attendance taking, using selected modern models. Three experiments conducted on the custom face detection dataset, LFW and CPLFW respectively, showed that RetinaFace was the best facial detection model while ArcFace with Cosine similarity and FaceNet512 with Euclidean L2 similarity were the best facial recognition models. Additionally, this project uncovered upscaling and ensemble embeddings as novel techniques that led to significant improvements in model performance. Using the best models and novel techniques analysed in the experiments, a prototype was developed. Evaluating the prototype on the custom evaluation dataset highlighted the limitations of current models and methods. The report thus concludes that large group facial recognition is still not feasible for practical uses as the prototype’s recognition accuracies, while decent, is still relatively far from the ideal accuracy of 100%.