Anil Kumar

Anil Kumar Vadathya

I am a master's student with Dr. Kaushik Mitra at IIT Madras, where I work on computational imaging, deep learning and image processing. At Computational Imaging lab , I've worked on deep learning based solutions for image restoration in compressive imaging setups like Single Pixel Camera, compressive light field reconstruction and RGBD priors. Prior to this I got my B. Tech in Electronics and Communications from RGUKT, Basar in 2015.

Our project titled "Deep Generative Models for Computational Imaging" received Qualcomm Innovation Fellowship (QInF), India 2016 and Super winner for the consecutive year 2017 among four prestigious projects.

Email / CV / Google Scholar

News:

(Nov 12th, 18) Our work on solving inverse problems using deep pixel-level prior is accepted at IEEE Trans. on Computational Imaging (TCI)!
(Nov 10th, 18) Joined as Research Engineer at Rice Univeristy, Houston, US!
Deep pixel-level prior for inverse computational imaging is accepted at TADGM workshop, ICML 2018! (June 15th, 18)
Poster on light field reconstruction from focus-defocus pair at CCD workshop, CVPR 2018! (May 30th, 18)
Poster demo on solving inverse problems using deep pixel-level prior at ICCP 2018, CMU, US! (Apr 25th, 18)
Our QInF 2016 got renewal to 2017! (among four prestigious projects)

Research

I'm interested in computational photography, image processing, optimization, and machine learning. Much of my recent work has focused on using deep generative modesl for inverse problems in computational imaging.

	A Deep Learning Framework for Light Field Reconstruction From Focus-Defocus Pair: a minimal hardware approach Anil Kumar Vadathya, Sharath Girish, Kaushik Mitra Computational Cameras and Displays (CCD) workshop, CVPR, 2018 poster / We provide a learning algorithm for light field reconstruction from a conventional camera.
	Solving Inverse Computational Imaging Problems using Deep Pixel-level Prior Akshat Dave, Anil Kumar V., Ramana Subramanyam, Rahul Baburajan, Kaushik Mitra Under review at IEEE Trans. on Computational Imaging Accepted at Deep Generative Models applications (TADGM) workshop at ICML, 2018! arxiv / code We use a deep autoregressive image prior PixelCNN++ for solving inverse imaging problems. Since autoregressive nature explicitly models pixel level dependencies it reconstruct pixel level details much better than existing state of the art methods.
	Learning Light Field Reconstruction from a Single Coded Image Anil Kumar V., Saikiran C., Gautham R., Vijayalakshmi Kanchana, Kaushik Mitra Asian Conference on Pattern Recognition, 2017 project page / poster Using deep neural networks we reconstruct full sensor resolution light field from a single coded image. Our approach involves depth based rendering where depth is learnt in an unsupervised manner.
	Compressive Image Recovery Using Recurrent Generative Model Akshat Dave, Anil Kumar Vadathya, Kaushik Mitra International Conference on Image Processsing (ICIP), 2017 project page / github / poster We use a deep recurrent image prior, RIDE, which can model long range dependencies in images very well. Using this for compressive image recovery we show much better reconstructions especially at lower measurement rates.
	Denoising High Density Gene Expression in Whole Mouse Brain Images Mayug Maniparambil, Anil Kumar V., Kannan U. V., Kaushik Mitra, Pavel Osten Society for Neuroscience (SfN), 2017 poster We use a deep autoencoder with adversarial loss for denoising the gene expression to improve the registration accuracy.
	Butterfly Communication Strategies: A prospect for soft-computing techniques Sowmya Charugundla, Anjumara Shaik, Chakravarthi Jada, Anil Kumar Vadathya International Joint Conference on Neural Networks (IJCNN), 2014 BMO code for three peaks function. We show mathematical models of communication mechanisms debloyed by butterflies. This work is extended to an optimization algorithm, Butterfly Mating Optimization (BMO), Jada et al. 2015. We applied it for image clustering application.
	ROBOG Autonomously Navigating Outdoor Robo-Guide Kranthi Kumar R., Irfan Feroz G. M., Chakravarthi Jada, Harish Y., Anil Kumar V. International Conference on Swarm, Evolutionary, and Memetic Computing, 2014 Neural networks are used to learn the navigation information. For outdoor navigation, we propose an image processing pipeline for road detection.