# Vijay Rengarajan

I have joined Image Science Lab headed by Prof. Aswin Sankaranarayanan at Carnegie Mellon University as a post doctoral associate in December 2017. Well, Pittsburgh is really cold, and I am not sure if I will get used to the unit of Fahrenheit.

I was a Ph.D. student at Indian Institute of Technology Madras. I worked on the study of camera motion in rolling shutter cameras with Prof. A.N.Rajagopalan and Prof. R.Aravind in Image Processing and Computer Vision lab.

Email: vangarai@andrew.cmu.edu

## Research problems

### Rolling Shutter Cameras

Most cameras of today use CMOS sensors, and almost all of which employ an electronic rolling shutter mechanism. The image exposure is not one-shot-global but row-wise. This mechanism affects traditional geometry analyses while working with such cameras. Any camera motion during exposure causes distortions even within an image. What you see with your human eyes is not what you get in the camera. Such distortions cause skew, wobble, curvature in the recorded image and the phenomenon is named as the rolling shutter effect.

Simulation Demo

Single Image Rectification Correcting rolling shutter effect from videos are addressed in most works in literature, while our work tackles the challenging problem of rolling shutter correction from only a single image. The challenges of this problem are: (i) identifying and extracting the distorted information, (ii) exploiting that information to estimate camera motion, and (iii) correcting the distorted image according to the estimated motion. We propose a geometry based method to estimate camera motion from curvatures of a distorted image of an urban scene Project Page, and also propose a convolutional neural network having long kernels to learn the distortion-to-motion mapping. Project Page

Super-resolution In classical super-resolution techniques, the warps between between low-resolution images are assumed to have arisen from global camera pose changes, and this is true as long as global-shutter sensors are used. When rolling-shutter sensors are amployed, we need to consider row-wise camera pose changes. Our framework registers rolling-shutter distorted low-resolution images row-wise in a subpixel manner and produces a distortion-free high-resolution image. Project Page

Image Registration and Change Detection This is a seminal work in the area of image registration for rolling shutter cameras. We perform row-wise image registration in the presence of both rolling shutter and motion blur artifacts. We work on the application of change detection. And therefore, in addition to geometric and photometric registrations, any changes between the images are also jointly estimated. The method is extremely useful in drone imaging when the imaging setup employs CMOS sensors. Project Page

### Global Shutter Cameras

Large-image Motion Blur Registration Change detection is an important task in aerial imagery in which environmental changes and new/lost/modified objects in the scene are detected by imaging from aircrafts over different periods of time. In such a scenario, registering large images both geometrically and photometrically in the presence of viewpoint change and motion blur is a challenging task. This work estimates camera pose differences between two large images quickly by using only a subimage of the whole image. Project Page

### Single Pixel Cameras

Camera Motion Estimation Through compressed sensing, we can beat Nyquist sampling rate by assuming sparse priors on the data to be captured. A single pixel camera does this for images by recording a limited number of linear combinations of the scene with random images. Each linear measurement is recorded one at a time. From such linear measurements, images can be got back through an optimization based on sparse priors on the image. It will work as long as the scene and the camera is stationary. Camera motion during such a sequential acquistion would be devastating since we will essentialy be looking to reconstruct different images of the same scene since the camera has moved during the acquisition. In this work, we estimate the camera motion difference between two such sensed measurements without actual image reconstruction that could be useful to remove these artifacts. Project Page