CCT: Consensual Collaborative Training and Knowledge Distillation based Facial Expression Recognition under Noisy Annotations
Arxhiv paper: https://arxiv.org/ftp/arxiv/papers/2107/2107.04746.pdf
This work has been accepted for publication in International Journal of Engineering Trends and Technology (IJETT), 2021.
Our proposed CCT framework
CCT involves co-training of three networks θ1, θ2 and θ3 jointly using a convex combination of supervision loss and consistency loss. Consensus is built by aligning the posterior distributions (shown as dotted red curves between p1, p2 and p3 ) using consistency loss. Dynamic weighing factor(λ) that balances both the losses is given by Gaussian like ramp-up function.
Authors: Darshan Gera and Dr. S. Balasubramanian, SSSIHL.
Any queries please mail to: darshangera@sssihl.edu.in.
Acknowledgments: We dedicate this work to Bhagawan Sri Sathya Sai Baba, Divine Founder Chancellor of Sri Sathya Sai Institute of Higher Learning, PrasanthiNilyam, A.P., India.