B.Tech(Hons.) in Computer Science
Center for Visual Information Technology
Department of Computer Science and Engineering
International Institute of Information Technology, Hyderabad

Office: CB #5132 Sitterson Hall, Chapel Hill, NC, USA
Residence: Chapel Hill, NC.
Email: ""firstname"" at students.iiit.ac.in
Phone Number:

 


Research

My broad research interests are Computer Vision, Computer Graphics and GPU programming.

I worked in geometric problems in computer vision, like large-scale structure-from-motion and robust estimation techniques. At IIIT-Hyderabad, I worked with Professor P.J.Narayanan.

After working for 2 wonderful years at Credit Suisse First Boston(CSFB) as an infrastructure developer, I'm now a student at University of North Carolina, Chapel Hill. New website - link


Projects


Undergraduate Thesis
    3D Reconstruction: Large scale structure from motion by reducing the computational bottleneck of the process.
    Shubham Gupta, Siddharth Choudhary and P.J.Narayanan.

    Ongoing undergraduate thesis, Under this project, the main stress is to reconstruct 3D models from 2 dimensional data (Photos) on the GPU.

    Publication Title: Practical Time Bundle Adjustment for 3D Reconstruction on GPU

    Abstract: Large-scale 3D reconstruction has received a lot of attention from the computer vision community recently. Bundle adjustment is a key component of the reconstruction pipeline. The bundle adjustment step requires a considerable amount of computational resources and is usually the slowest step in the pipeline. This step hasn't been parallelized effectively either. In this paper, we present a hybrid implementation of sparse bundle adjustment on the GPUs using the CUDA programming model, with the CPU working in parallel. The overall algorithm is decomposed into smaller steps. Each of which is scheduled on the GPU or the CPU. We develop efficient kernels most of the steps and exploit existing libraries for data parallel operations and matrix inverse. Our implementation outperforms the CPU implementation significantly, achieving a speedup of 8-10 times over the standard CPU implementation for datasets with upto 500 images on one quarter of an Nvidia Tesla S1070 GPU.

    In the Proc. of ECCV 2010 Workshop on Computer Vision on GPUs [pdf] [ppt]


Last updated: Oct 2012