Avirup Dey

I am a 2nd-year Computer Science PhD student at the University of Bath, advised by Prof. Vinay Namboodiri. I work on controllable generative models for 3D facial animation.

I completed my B.E. in Electronics and Communication Engineering in 2023, from Jadavpur University, India.
Previously, I was an undergraduate research assistant at DVLP-CMATER Lab, Jadavpur University under Prof. Nibaran Das, where I worked on deep learning-based models for document analysis and cell imaging. As a recipient of the Mitacs Globalink Research Internship, I spent the Summer of 2022 at Ontario Tech University, where I worked under Prof. Mehran Ebrahimi on MRI Super-Resolution. I have also worked as a research assistant at the IVPR Group under Prof. Ananda Shankar Chowdhury, where I worked on domain adaptation methods and event cameras.

Email  /  CV  /  Twitter  /  Scholar  /  Github

profile photo
Research

I'm broadly interested in controllable generative models. Currently, I'm working on steering diffusion models for facial animation. During the initial months of my PhD, I worked on audio-driven talking faces with 3D Gaussian Splatting in collaboration with DeepReel. Have some fun watching Elon imitating Marcus Antonius!
As an undergrad, I worked on inverse inverse imaging tasks like super-resolution and binarization. Check out my past works below.

MTSR-MRI: Combined Modality Translation and Super-Resolution of Magnetic Resonance Images
Avirup Dey, Mehran Ebrahimi
MIDL, 2023
code / paper

TLDR: Pilot study on cross-modal MRI super-resolution. Optimized off-the-shelf i2i models to leverage sparsity of the scans.

Variational Augmentation for Enhancing Historical Document Image Binarization
Avirup Dey, Mita Nasipuri, Nibaran Das
ICVGIP, 2022
code / paper

TLDR: GAN-based augmentation helps segmentation models perform better on document binarization benchmarks (DIBCO).

ResCNN: An Alternative Implementation of Convolutional Neural Networks
Avirup Dey, Sarosij Bose
IEEE UPCON, 2021
code / paper / slides

TLDR: Expressing kernel convolution as a matrix-matrix product and reducing image rank boosts training efficiency.

Misc
1. Won AI-Entrepre-Neural 2021 organised by IIT Kharagpur
2. Hosted a webinar with GATE 2022 toppers.
3. Hosted Convolve 1.0, a virtual machine learning workshop for our freshmen.

© Avirup Dey. Template borrowed from John Barron