I am interested in problems related to deep learning, computational photography and computer vision.

Deep Learning:

Deep learning has revolutionized applications in CV and ML. I am interested in supervised as well as unsupervised learning algorithms. My recent work in this area includes building deep learning models for semantic segmentation and unsupervised learning algorithm for recommending complementary items.

Context Encoding for Semantic Segmentation, CVPR 2018 (oral presentation)

CRAFT: Complementary Recommendations Using Adversarial Feature Transformer

Computational Photography:

With advances in sensors and computational power, cameras coupled with a computer offer us new possibilities which were not possible with traditional film based cameras. Computational photography is emerging as a new field combining computer vision, graphics and imaging to overcome the limitations of current cameras. Traditional approaches try to combine multiple photos with varying parameters to overcome the limitations of the cameras. However, there is a need for novel sensors for specific applications that go beyond the traditional capturing of the scene as a regular grid of pixel intensities. Such sensors, for example, include coding and modulation strategies along dimensions of space, time, angle and/or wavelength.

I have creatively devised novel solutions for many classic imaging problems like motion blur, narrow depth of field, low frame rate, lens glare, and photo artifacts due to flash and glass reflections. These problems are manifestations of inevitable tradeoffs and loss of visual information in photography. My emphases have been on analyzing the underlying loss of information and modifying the image capture process itself to solve the problem, in contrast to traditional software-only approaches which are often inadequate. In solving these problems, my goal is to design easy, low cost solutions that anyone can build as well as to enhance off-the-shelf imaging devices.

Novel Sensors and Imaging Devices

Sensor/Imaging Device


Coding/Modulation Dimensions

Flutter Shutter Camera (Coded Exposure)
CVPR 2007, CVPR 2009
Time (Within an Exposure)
Video Motion Deblurring
SIGGRAPH 2009 Time (Across Exposures)
Coded Camera Array for Temporal Super-Resolution CVPR 2010
Time (Across Cameras)
Gradient Camera
CVPR 2005
Intensity Values
Coded Aperture Camera
Mask based Heterodyne Light Field Camera
SIGGRAPH Asia 2008,
Space and Angle
Reinterpretable Imager
Eurographics 2010
Space, Time and Angle
Per-Pixel Shutter for Flexible Space Time Sampling
ECCV 2010
Space & Time
Wide Angle Light Field Camera using Mirror Arrays & Crystal Balls
SIGGRAPH Asia 2010
Space and Angle (wide FOV)

Physics based Modeling:

Modeling and understanding specular reflection is important for several vision tasks such as reconstruction of specular surfaces, recognition and pose estimation of specular objects as well as catadioptric sensors that use a mirror to obtain wide angle photos. Although catadioptric sensors have been widely used for last two decades, most of the focus has been on using central catadioptric systems, which can be easily modeled.

Beyond Alhazen's Problem: Analytical Projection Model for Non-Central Catadioptric Cameras with Quadric Mirrors, CVPR 2011

A. Agrawal, Y. Taguchi, & S. Ramalingam, "Analytical Forward Projection for Axial Non-Central Dioptric and Catadioptric Cameras", ECCV 2010

Y. Taguchi, A. Agrawal, S. Ramalingam & A. Veeraraghavan, "Axial Light Fields for Curved Mirrors: Reflect Your Perspective, Widen Your View", CVPR 2010

Y. Taguchi, A. Agrawal, A. Veeraraghavan, S. Ramalingam and R. Raskar, "Axial-Cones: Modeling Spherical Catadioptric Cameras for Wide-Angle Light Field Rendering", SIGGRAPH Asia 2010

A. Sankarnarayanan, A. Veeraraghavan, O. Tuzel, & A. Agrawal, "Image Invariants for Smooth Reflective Surfaces", ECCV 2010

A. Sankarnarayanan, A. Veeraraghavan, O. Tuzel & A. Agrawal, "Specular Surface Reconstruction using Sparse Reflection Correspondences", CVPR 2010

Gradient Domain Algorithms:

Before Deep Learning, gradient domain algorithms have become popular in computer graphics for applications such as image editing, matting, HDR compression, video surrealism and context enhancement. My previous work in this area novel gradient based methods for "visiony" problems such as (a) recovering intrinsic image (separating illumination and reflectance) (b) removing shadows from color images (c) Edge suppression under significant illumination variations (d) removing glass reflections and (e) surface reconstruction from gradient fields.

Relevant Publications:

A. Agrawal,  R. Raskar,  Shree K. Nayar &  Y. Li,  "Removing Photography Artifacts using Gradient Projection and Flash-Exposure Sampling",  ACM Transactions on Graphics (Proceedings of  SIGGRAPH) 2005  

A. Agrawal, R. Raskar and R. Chellappa, "Edge Suppression by Gradient Field Transformation Using Cross Projection Tensors", IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2006 pdf    Matlab Code

A. Agrawal, R. Raskar and R. Chellappa, "What is the Range of Surface Reconstructions from a Gradient Field?", European Conference on Computer Vision (ECCV), 2006  (oral presentation, 4.5% acceptance)  pdf    Matlab code GUI    Matlab code Mfiles

A. Agrawal, R. Chellappa & R. Raskar, "An Algebraic Approach to Surface Reconstruction from Gradient Fields",  IEEE International Conference on Computer Vision (ICCV), 2005 pdf   Matlab code

Phd Thesis, "Scene Analysis under Variable Illumination using Gradient Domain Methods", Dept. of Electrical and Computer Engineering, University of Maryland, College Park, 2006