Publications
Anup Shetty, Sumantra Dutta Roy, Subhasis Chaudhuri, "Importance Sampling-Based Probabilistic Eigentracker", Proc. National Conf. Computer Vision, Pattern Recognition, Image Processing and Graphics, Gandhinagar, India, Jan 2008.
Research Projects
Computer Vision based Human Activity Analysis
Supervisor: Dr. Subhasis Chaudhuri,
V&IP Lab, Department of Electrical Engineering, IIT Bombay
In this project we implemented different clustering/classification based and model based algorithms to recognize different human activities like running, skipping, waving etc. Variants of K-means and C-means algorithms were implemented. Hidden Markov Model and Conditional random fields were also used to recognize the activities. These algorithms were also tested on database containing different human-human interaction. It was found that the model based approaches gives better accuracy over the clustering/classification methods.
Particle Filtering based EigenTracker
Supervisor: Dr. Sumantra Dutta Roy
SPANN Lab, Department of Electrical Engineering, IIT Bombay
Markov chain Monte Carlo based moving object tracker was implemented. Both motion and colour cues were used for initialization. Constant acceleration model was used for handling acceleration. The appearance information was calculated using Black and Jepson's Eigenspace representation. Particle filtering algorithm with Importance sampling was used where the affine parameters of the bounding box of the tracked object were taken as particles.
Image compression using Set Partitioning In Hierarchical Trees
Supervisor: Mr. Jayram Kelawade
Department of Electronics Engineering, DMCoE
Set Partitioning In Hierarchical Trees (SPIHT) is a wavelet based coding algorithm that exploits the inherent similarities across the subbands in a wavelet decomposition of an image. This algorithm codes the most inmportant wavelet transform coefficients first, followed by the detail coeficient. SPIHT was implemented in Matlab and was tested for different types of images. The results were also compared with JPEG and naive wavelet based compression algorithms.
Graduate Course Projects
Variant of JPEG implementation
Implemented a variant of JPEG where Lempel-Ziv coding was used instead of run length coding
Image Processing
Developed image processing library in C++ with support for basic image manipulations, Fourier transform, filtering, feature detection, mosaicing, morphing
Computer Vision
Implemented Vision based algorithms like Shape from Shading, Optical Flow and Depth from Stereo
SVD based Digital Watermarking
Implemented Singular Value Decomposition based Digital Image Watermarking technique
Was awarded best project among 20 other groups
Speech Processing
Implemented isolated digit recognizer using Continuous Random Fields using Mel-Frequency Cepstral Coefficients as features
Adaptive Digital Beamformer
Implemented adaptive algorithms like LMS, RLS for Digital Beamformer
Tested the algorithm for various sensor orientations