Project Computer Science

Siddharth Choudhary

Centre for Visual Information Technology

Department of Computer Science and Engineering Ph:+91-994*-***-***

IIIT Hyderabad http://web.iiit.ac.in/ siddharth ch

OBH 134, IIIT Hyderabad, Hyderabad 500032 email: *********.*********@********.****.**.**

Education

IIIT Hyderabad, India 2010 - Present

MS (By Research) in Computer Science and Engineering

Advisor: Prof. P J Narayanan

IIIT Hyderabad, India 2006 - 2010

Bachelor of Technology in Computer Science and Engineering

(GPA: 8.5/10.0)

Publications

Siddharth Choudhary, Shubham Gupta and P J Narayanan. Practical Time Bundle

Adjustment for 3D Reconstruction on GPU, In Proceedings of the ECCV Workshop on

Computer Vision on GPUs (CVGPU 2010), Crete, Greece, Septemeber 2010

Presented the intial work on Bundle Adjustment in the Indian Conference for Academic Research by

Undergraduate Students (ICARUS 2010), Kanpur, India, March 2010

Experience

Research Assistant at CVIT, IIIT Hyderabad 2010 - Present

Ongoing work on large scale structure from motion by reducing the computational bottleneck of the

process

Software Develpment Intern at DrishtiCare Inc Summer 2010

Worked on the implementation of medical image analysis algorithms for the detection of Diabetic

Retinopathy.

Summer Intern at IIIT Hyderabad Summer 2009

Worked on the implementation of computer graphics algorithms for realtime rendering of gems.

Academic Achievements

Recipient of Undergraduate Research Award for the year 2009-2010 and 2010-2011.

Included in the Dean s List for the year 2007-2008 at IIIT Hyderabad.

Showcased the research work at the R&D Showcase 2009, a showcase of all major research

activities being done at IIIT Hyderabad.

Secured All India Rank 1070 among more than 500,000 candidates in All India Engineering

Entrance Examination 2006.

Major Projects

Fast Localization and Visualization of Personal Photo Collection using Community

Photo Collection (Ongoing)

In this project, we explore methods for fast localization and visualization of photo collection

captured mostly using a single camera over a period of daytime. We use the reconstructed point

cloud from community photo colleciton as a model for localization of new query image.

Fast and Scalable sparse non linear least squares optimization techniques in Geometric

Vision

Sparse non linear least squares optimization acts as a bottleneck in many of the vision algorithms. In

this project we explore methods to exploit the sparsity of the underlying data and the use of

multicore architectures in order to make the whole optimization fast and scalable. We explore the

opportunity to use multiple GPUs along with multicore CPU inorder to solve large scale Bundle

Adjustment for faster 3D reconstruction.

Large scale structure from motion by reducing the computational bottleneck of the

process (Undergraduate Thesis)

Large scale Structure from Motion has recieved a lot of attention in the computer vision community

with applications in virtual photo tourism and city scale modelling. In this project, the main aim

was to identify the computational bottlenecks in the large scale 3D reconstruction process and to

solve them using faster algorithms with the help of multicore architectures.

Analysis of the Performance of Various Loss Functions in SVM

Recent studies about non-convex loss functions have shown them to be faster and more able to

handle large scale problems than convex functions. The CCCP procedure used to handle these

optimization problems guarantees convergence towards the solution with each iteration. The ramp

loss function, in particular, has been shown to give lesser number of SV for a given problem than the

traditional hinge loss function. This prompts towards the analysis of various other loss functions.

The loss functions which saturate or decrease for large distances from the margin have an advantage

that they do not incur high penalties for incorporating noise points and hence possibly can handle

very noisy datasets better.

Image Fusion for Context Enhancement

This is a method based on gradient domain techniques to fuse di erent images of the same scene

under di erent illumination condition to enhance low quality night time image. This is a

implementation of the paper published in 2004 ACM. Done as a course project for Digital Image

Processing.

Multi GPU Implementation of SIFT Detector

Scale-invariant Feature Transform algorithm is a powerful algorithm to extract information from a

real-world image. This is an implementation of the detector part of this algorithm, which works on

Single/Multi-GPU machine.

Rotation Invariant Object Recognition from one Training Example

This is a robust method of object recognition based through local descriptors based on Gaussian

Derivatives. This is an implementation of report published by CSAIL, MIT. It was done as the

course project of Pattern Recognition.

Skill Set

Operating Systems: Windows, Linux

Programming Languages: C, C++

Scripting Languages: Bash, Python, PHP, Matlab

Libraries and APIs: OpenGL, OpenCV, CUDA, OSG, MySQL

Web-Based Technologies: HTML, CSS, Javascript, AJAX

Programming Environments: Vim, Visual Studio

Conferences Attended

European Conference on Computer Vision 2010, Crete, Greece

Indian Conference on Computer Vision, Graphics and Image Processing 2010, Chennai, India

Indian Conference on Academic Research by Undergraduate Students 2010, Kanpur India

Relevant Courses Undertaken

Data Structures, Programming Languages,Graph Theory, Computer Vision, Machine Learning, Computer

Organization, Digital Image Processing, Computer Graphics, Pattern Recognition, Arti cial Intelligence,

Database Management System, Spatial Informatics, Speech Technology

Extracurricular Activities

Avid interest in programming competitions.

Website Designer and Developer of Felicity (College Fest) Website.

Member of the core technical team of the college fest.

Contact this candidate