Rakshit Gautam

ACM Recys Challenge 2013 : Yelp Business Rating Prediction
Guided by Dr. Ernesto Diaz-Aviles at L3S Research Center, Hannover, Germany
- Created a recommender system for rating prediction on Yelp dataset
- Applied collborative filtering techniques to generate models and ensembled them to achieve 3% RMSE decrease over global user mean baseline
- Dealt with problems of sparse data and cold start

Driverless Vehicle: Mahindra, Spark the Rise Challenge ( Automotive Research Team: Cube26 Pvt. Ltd , New Delhi )
- Developed packages for classification of pedestrians, bicycles and cars from live street image data, on Robotics Operating System (ROS) framework
- Computed classifiers with SVM-Light using Histogram of Oriented Gradient (HOG) features

Power Test automation and analysis (Samsung R&D Institute India, Noida)
- Developed an android application (Sysresource) that provides real time system load statistics
- Used Monkey Runner and Android View Client to automate power tests used for mobile testing

Facial Expression Recognition
Worked under the supervision of Prof. KK Biswas at IIT Delhi. I developed a real time expression recognition system that is robust to illumination changes. I did a literature survey of several techniques used in the domain, and tried to improve upon the state of the art. I computed features using Local Binary Patterns, Local Directional Patterns, Local Binary Patterns in three orthogonal planes and geometric displacement of interest points obtained from a shape model. I implemented Naive Bayes, SVM with kernels and Nearest Neighbor algorithms to create expression classifiers. I also tried weber normalization of face images and divided the face into varying sized windows for feature computation. The models were trained and validated on a sizable chunk of data from FEED and Cohn-Kanade datasets. Results showed that weber normalization of face images results in decreased information content, and moderate window sizes best capture the expression information. The final model generated was fairly accurate on real world data.

Functional Connectivity Utility for AFNI
Worked under the supervision of Prof. Rahul Garg at IIT Delhi. I developed a system that facilitates the interactive analysis of functional MRI data on AFNI. The system asks user to choose a region of interest (ROI), method to generate representative time series for ROI, the correlation method to compute correlation and displays correlation maps to glean statistical insights into data.
Demo