Decoding cognitive states from brain fMRI scans
With Chittibabu, under the guidance of
Dr Krithika Venkataramani
In this project, we investigated whether it is possible to decode what a person is thinking, analyzing his/her brain fMRI
scans. Objective was to employ machine learning techniques on fMRI data, so that one could predict whether the person is
seeing a picture, or reading a sentence. Since seeing and reading are associated with completely different parts of the brain,
its theoretically possible to differentiate one from other. The biggest issue that we had to deal with was the highly hyper-dimensional dataset.
fMRI scans typically lies in ~100,000 dimensional space. We proposed a new approach for dimensionality reduction, using which our results were
significantly more accurate than the current state of the art method.
Results
Links
Project report
Presentation
fMRI dataset
Rotation invariant face detection based on Real Adaboost
Under the guidance of
Dr Amitabha Mukerjee
Objective of this project was to develop a high performance face detection system, which works fine irrespective of the facial orientation.
The popular Viola-Jones algorithm suffers from its inability to handle face images that are in different poses. To overcome this problem, we augmented
the Viola-Jones algorithm with real adaboost algorithm, instead of the traditional discrete adaboost. This helped us to parallelize multiple
detection cascades to gather, each one trained for a specific orientation of the face. Real adaboost gave us an additional opportunity to reduce the computation,
by introducing a ranking scheme to rank the parallel cascades, so that after a few stages, one can stop the cascades with low rankings.
Results
Links
Project report
Presentation
PIE dataset
Constructing 3D face models from depth and RGB data
Under the guidance of
Dr Pratwijith Guha
This project was done as a part of my summer internship at TCS Labs, Delhi. We hacked Microsoft Kinect with the help of Point Cloud Library, so that it can be used
to build 3D human face models. Firstly, we combined the RGB data with the depth data captured by the infrared camera in Kinect , to form a 3D model of the environment. Then
a face detection algorithm was ran on the RGB stream so that the corresponding 3D face in the point cloud was able to extract. Now, having the partial 3D face clouds, next
job was to stitch them together to form a complete model of the face. This was a 2 phase procedure. An initial alignment was done based on the Point feature histogram, followed
by an optimization step that used a variant of Iterative Closest Point algorithm to minimize the alignment error. Once a 3D point cloud was obtained, a ball-pivoting
surface reconstruction algorithm was applied get the mesh model of the face.
Results
Links
Presentation
Adaptive fingerprint enhancement and cross-matching
With Muralidharan,
Mithun NK,
Vijith KK,
Saritha Murali,
under the guidance of Smithamol B
This was done as a part of my B tech project, and as a part of this project, we developed a new algorithm for fingerprint matching, which have high tolerance against
smudged/dry fingerprints. We used a neural network preprocessing stage, which processed fingerprints adaptively, based on its nature[dry/normal/oily]. After that,
we used a novel "distance vector" algorithm that we developed to match the fingerprints. Our paper on the same was accepted in three international IEEE conferences, namely ICACTE
2010, ICCSIT 2010 and ICUMT 2009. It was ranked among the top 10%-30% of all the papers in ICUMT 2009 .
Links
Project Report