Trunk-Branch Ensemble Convolutional Neural Networks for Large Scale, Few-Shot Video-to-Still Face Recognition

Trunk-Branch Ensemble Convolutional Neural Networks for Large Scale, Few-Shot Video-to-Still Face Recognition

Chen, Joe (TU Delft Electrical Engineering, Mathematics and Computer Science)

Yorke-Smith, Neil (mentor)
Mocking, Daniël (graduation committee)
Urbano, Julián (graduation committee)

Delft University of Technology

Computer Science | Data Science and Technology

2019-07-29

For the real-world face recognition, factors such as occlusion and pose-variant (cross face) would affect the identification/verification performance. In addition, large number of classes also increase the complexity, which makes verification/identification even harder. In order to deal with these issues, how to extract discriminative embeddings is a challenging task for the researchers. This research aims at video-to-still (V2S) face identification, which means given few images per person (p.p.) as our database (also called gallery), we tend to identify if a person in a video is someone in our database or not. We use end-to-end Trunk-Branch Ensemble Convolutional Neural Networks (TBE-CNN) combined with state-of-the-art InceptionNet to create informative image patches and boost the features for occlusion and pose-variant issues. The images for learning and identifying comprise occlusions, different face directions and also resolution issues. Moreover, we create a large scale, few shot video-to-still (still-to-still) face recognition dataset with different settings to evaluate the the models and find the preferred settings for real-world face recognition application. How to ensure the accuracy under these practical noises and settings is the goal of this research.

Face Recognition
Convolutional Neural Networks
Trunk-Branch Ensemble Convolutional Neural Networks
Occlusion
Pose-variance
Real-world application
Video-to-still Face Recognition
Still-to-still Face Recognition

http://resolver.tudelft.nl/uuid:575b2bcb-46cb-471e-b1cb-565fbe125cdb

Student theses

master thesis

© 2019 Joe Chen