Projects done

---

CoinRun - Game AI

Spring 2019

Guide: Prof. Mark Riedl, College of Computing, Georgia Tech

• Implemented a Deep Reinforcement Learning agent using PyTorch that plays a platform game CoinRun
• Trained a Deep Q Network which predicts the best possible action based on a reward function

Summer 2018

Guide: Prof. Amitabh Varshney, Maryland Blended Reality Center, University of Maryland

• Implemented a parallel version of the the Quadric Error Metric method to perform Mesh Simplification on GPU using CUDA
• Simplified the mesh in Real Time in under 100ms compared to the traditional 700ms for a serial implementation

Fall 2017

Guide: Prof. Jarek Rossignac, College of Computing, Georgia Tech

• Computed the Delauney Tetrahedralization of two given clouds of balls located at two horizontal planes
• Computed a high-resolution water-tight triangle mesh that approximates the boundary of the union of balls
• Detailed Report   Source Code

Fall 2017

Guide: Prof. Jeff Wilson, Interactive Media Technology Center, Georgia Tech

• Created a 3D single-player racing game in Unity
• Created a player character with animated 3D mesh character controller having real-time control
• Implemented a real-time steering, path planning and state-machine based AI which controls 3 NPCs
• Added physics event-based feedbacks like particle effects and 3D audio
• Final Presentation   Source Code

Fall 2017

Guide: Prof. James Hays, College of Computing, Georgia Tech

• Loaded cropped positive trained examples (faces) and converted them into a HoG template
• Sampled random negative examples (non-faces)
• Trained a classifier from these examples
• Ran the classifier on the test set and achieved 82.5% average precision
• Detailed Report   Source Code

Fall 2017

Guide: Prof. James Hays, College of Computing, Georgia Tech

• Classified scenes into one of 15 categories by training and testing on a 15 scene database
• Used features like tiny images and SIFT and classifiers like nearest neighbor and linear SVM
• Also experimented with RBF kernel for non-linear SVM, GIST descriptors and Fischer encoding
• Achieved best accuracy of 65.5% using SIFT and linear SVM
• Detailed Report   Source Code

Fall 2017

Guide: Prof. James Hays, College of Computing, Georgia Tech

• Calculated the camera projection matrix and camera centers for given 2D and 3D points of an image
• Estimated the Fundamental matrix for 2 input images using Singular Value Decomposition
• Used RANSAC to find the best Fundamental matrix for matches of 2 images found using SIFT
• Detailed Report   Source Code

Fall 2017

Guide: Prof. James Hays, College of Computing, Georgia Tech

• Implemented the Harris Corner Detection algorithm to find interest points of two similar images
• Implemented a SIFT like feature detection algorithm to find local features for each of the interest points
• Used Nearest Neighbour Distance ratios to find the feature matches between the two images
• Detailed Report   Source Code

Fall 2017

Guide: Prof. James Hays, College of Computing, Georgia Tech

• Created a filtering algorithm which applies a filter to a given image
• Created low pass and high pass filters of two different images and merged them to create a hybrid image
• Detailed Report   Source Code

Spring 2017

Guide: Prof. Siddhartha Chaudhuri, CSE Dept, IIT Bombay

• Created a parser for a grammar of a city and parsed it to create a syntax tree
• Iterated over the faces of a manually generated road network and called a render function at each leaf node
• Probabilistically generated different types of buildings like schools, offices, residential homes etc
• Coded it in C++ using graphics libraries like OpenGL and GLUT

Spring 2017

Guide: Prof. Ajit Rajwade, IIT Bombay

• Designed a system for real-time tracking of non-rigid objects from a moving camera using Mean Shift
• Used Bhattacharyya Coefficient based metric for better target localization

Autumn 2016

Guide: Prof. Parag Chaudhuri & Siddhartha Chaudhuri, CSE Dept, IIT Bombay

• Designed a system that scans a human body using multiple Depth Sensors like Microsoft Kinect
• Robustly reconstructed a synthetic mesh of a person using these partial, noisy scans
• Detailed Report   Presentation

2016

Guide: Prof. Pushpak Bhattacharyya, IIT Bombay

• Studied and Compared various feature selection methods like TFIDF, Delta-TFIDF, Relief, Χ² test and Welch’s t-test
• Performed In-domain, Cross-domain and Cross-lingual SA in and across various domains/languages
• Showed that t-test achieves higher accuracy than Χ² test in In-domain, Cross-domain and Cross-lingual SA
• Detailed Report   Presentation

Summer 2016

Guide: Pradeep Dwarakanath, Sr. Chief Engg., Samsung R & D Institute Bangalore

• Studied the behavior of secondary SIM in case of switching from one SIM to another in Dual SIM phones
• Used various probabilistic models to learn and predict the behavior of the secondary SIM
• Employed smart scheduling strategies at network to minimize the loss of "On Air" resources
• Tested the code with multiple configurations and showed improvement in resource utilization at NW

Spring 2016

Guide: Prof. Parag Chaudhari, IIT Bombay

• Created a low-cost facial animation system using a non-intrusive, commercially available 3D sensor (Kinect)
• Reconstructed the set of user-specific blendshapes that best reproduce the example expressions
• Represented dynamics of facial expressions using a generic blendshape rig
• Coded it in C++ using PCL tools and library to generate and store point clouds
• Detailed Report   Source Code

Spring 2016

Guide: Prof. Parag Chaudhari, IIT Bombay

• Designed a humanoid (our own creation) and a non-humanoid (BB-8) bot, inspired from the Star Wars movies
• Used multiple point lights which acted as an area light and generated soft shadows
• Used indirect illumination for Color Bleeding and Photon Mapping for Caustics
• Coded it in RSL and rendered in RenderMan, a renderer by Pixar
• Detailed Report   Source Code

Spring 2016

Guide: Prof. Amitabha Sanyal, IIT Bombay

• Implemented a compiler for a subset of the C language using tools like FlexC++ and BisonC++ askk
• Source Code

Autumn 2015

Guide: Prof. Parag Chaudhari, IIT Bombay

• Designed a humanoid and a non-humanoid (BB-8) bot, inspired from the Star Wars movies
• Textured the bots and surrounding environment and added light sources for the background
• Added features like constrained rotation and translation of every part of the bots along 3 specified axes
• Coded it in C++ using graphics libraries like OpenGL, GLFW and GLEW
• Detailed Report   Source Code

Autumn 2015

Guide: Prof. N.L. Sarda, IIT Bombay

• Built a Java based web portal where people in an organization (like IIT Bombay) can Buy and Sell stuff they want
• Used HTML5, JSP, Java, CSS, Javascript for designing the web pages and PostgreSQL for handling the database
• Added features like profiles, messaging, searching and filtering of items
• Detailed Report   Source Code

Autumn 2015

Guide: Prof. Siva Kumar G., IIT Bombay

• Developed a Music Genre Classifier using Neural Networks
• Tested the classifier on various kinds of inputs for classifying them into pop, classical, metal, rock etc
• TStudied different parameters like total error, sensitivity and specificity and achieved > 80% accuracy
• Coded it in Python using libraries like Pybrain and Neurolab
• Detailed Report   Source Code

Autumn 2015

Guide: Prof. N.L. Sarda, IIT Bombay

• Implemented External Merge-Sort in PF Layer and Bulk Loading in AM Layer in B+ Trees
• Analyzed different parameters like time to construct the tree and number of nodes in the tree
• Studied the parameters in both Top-Down and Bottom-Up approach using pre-sorted and random data
• Detailed Report   Source Code

Summer 2015

Guide: Prof. Sharat Chandran, IIT Bombay

• Built a Django based web portal where JEE Mains/Advanced students can enter their ranks and interests
• Based on a statistical model, checked the available branches with different probabilities
• Used HTML5, Bootstrap, Javascript, jQuery for designing all the web pages
• Developed a feature using which students can directly upload final choices on official Seat Allocation portal

Summer 2015

Guide: Prof. Sharat Chandran, IIT Bombay

• Designed and implemented an algorithm to automate the Branch Change process
• The candidates are taken in CPI order and are assigned a branch based on predefined rules
• Implemented the algorithm using Java, it promises best possible programme to top candidate
• Tested and verified the algorithm on the actual dataset constructed from last two years
• Being used for the Branch change program in 2015 by the institute which was done manually till 2014

Spring 2015

Guide: Prof. Kameswari Chebrolu, IIT Bombay

• Implemented a chat system for both TCP and UDP protocols where we use the client to chat with a server
• Coded it in C++ using networking libraries like socket.h, arpa/inet.h etc

Spring 2015

Guide: Prof. Ashutosh Trivedi, IIT Bombay

• Designed a UART to receive/send data to microprocessor through data bus from/to a laptop
• Implemented it using VHDL and used Tera Term to transmit data frame by frame across FPGA and Computer

Autumn 2014

Guide: Prof. Sharat Chandran, IIT Bombay

• Developed a Python based web portal using Django Framework which included registration, login and session handling
• Implemented features like past year data analysis and available branches.

Autumn 2014

Guide: Prof. Sharat Chandran, IIT Bombay

• Studied the Gale Shapley Algorithm for finding a perfect matching and implemented it in Java.
• Used the implementation of the algorithm to solve a practical seat allocation based problem.

Autumn 2014

Guide: Prof. Sharat Chandran, IIT Bombay

• Modeled a simulation of a Rube Goldberg’s Machine using Box 2D , a physics simulation engine, in C ++
• Designed a robust system to develop a simulation using objects like dominos etc

Spring 2014

Guide: Prof. R.K. Joshi, IIT Bombay

• Designed a game where different objects of changeable shapes fall from the top
• Implemented it in C++ using a cross-platform graphics library - Fast Light Toolkit (FLTK)
• Objective is to make complete lines at the bottom. Implemented a scorebox and future object prediction too

Spring 2014

Guide: Prof. R.K. Joshi, IIT Bombay

• Designed a game where the player needs to reach the final destination from a starting point in a maze
• Implemented it in C++ using a cross-platform graphics library - Fast Light Toolkit (FLTK)
• Made it in two versions: one having hidden and the other having visible obstacles
• It takes inputs from arrow keys and the hidden obstacle version also shows the nearby available paths in different colors

Autumn 2013

Guide: Prof. Supratim Biswas, IIT Bombay

• Designed a game where the goal is to move an object to the top using jump and arrow keys to move sideways
• Implemented it in C++ using a graphics library called Simplecpp
• Added a feature in which the doodle (or object) changes shape and color when it touches the plank