Do Humans Fixate on Interest Points ?
Nanyang Technological UniversityKAUSTADSC, Singapore

In the field of Computer Vision, popular Interest Point detectors such as SIFT, SURF, MSER etc. have been used for a number of high level tasks including object detection, object recognition and image thumbnailing. This project primarily explores the question : What is the perceptual relevance of these interest points, in other words, how well do these correlate with human fixations on an image? In the first of its kind quantitative study (published in the proceedings of IEEE ICPR 2012), we found that the correlation is significantly weak. This project was done under the guidance of Dr. Bernard Ghanem. For more details Please refer to Do Humans Fixate on Interest Points?

Predicting Human Fixation Regions
Nanyang Technological University – KAUST – ADSC, Singapore

Having discovered that there exists a weak correlation between human fixations and detected interest points and on an image (Do Humans Fixate on Interest Points?), the results seem to indicate that there may be merit in learning from actual human fixations. Such a learning machine might be useful to assist current high-level computer vision tasks (more accurately) . As a continuation of the previous project, we created a model for learning and predicting regions containing human fixations in images.  The details of the method and results (over 85% positive patch detection rate and over 80% true negative elimination rate) have been recently submitted to the IEEE Transactions on Image Processing. Applications to Object Detection, Saliency Enhancement and Image Thumbnailing have also been suggested.

Nanobiosensors modeling and applications to early cancer detection [Project-Page]
LSI, École Polytechnique Fédérale de Lausanne, Switzerland

The missing circuit element – the memristor – has recently been prototyped ! These incredible devices have the property of holding state even when they are “powered-off”. Can these be used to detect changes in antibodies (disease fighting cells) in the human body? Can these help detect early forms of Cancer? These were the questions we posed.

Under the guidance of Prof. Sandro Carrara, along with Davide Sacchetto, I studied and modeled the current-voltage characteristics of antibody-functionalized nanowires (which show the memristive effect). A number of experiments were carried out to test the response to varying antigen concentrations under controlled physical conditions. Though still not fully understood, the model was able to very closely mimic the real device behaviour. Our Silicon Nanowire memristor could sense very low concentrations (femto Molars) of antigen change. This could seriously mean a step towards detecting very low concentrations of a specific antibody produced in specific forms of cancer. For more details, please see the journal paper : “Memristive Biosensors Under Varying Humidity Conditions“, F. Puppo, A. Dave, M.-A. Doucey, D. Sacchetto, C. Baj-Rossi, Y. Leblebici, S. Carrara, and G. De Micheli, IEEE Transactions on NanoBioscience.

Unsupervised Feature Selection
Nanyang Technological University, Singapore

When the number of dimensions of data becomes too large and computationally infeasible, we usually rely on techniques such as Principal Component Analysis for dimensionality reduction. However, if we wish to preserve the original dimensions of the data, these methods may not be apt. This project explores Information Theoretic and clustering based methods coupled with classic feature selection algorithms such as Relief. Applications include understanding of differential expression of Breast Cancer gene data and cancer progression data. This project was done under the guidance of Prof. Manoranjan Dash, Nanyang Technological University.

Low-Cost Geo-Sensors
Nanyang Technological University, Singapore

Final Year Thesis on “Explorations in the design of a multisensor system for seismic and geodetic monitoring of volcanoes” with Dr. Steven Wong and Dr. Dannie Hidayat (Earth Observatory of Singapore). The abstract can be found at the Nanyang Technological University repository. Find the full story at Multisensors for geological sensing.

The Maze Exploring Micro-mouse
Nanyang Technological University, Singapore

A fine example of Systems Engineering, this little mouse is adept at following black lines and detecting obstacles. When coupled with its control software (Java Master), it can explore mazes with obstacles and  report the shortest path. This project was done in a team of 6 with me as the lead.  The hardware prototyping, software engineering, maze exploration algorithm design and finally the integration were all done within a span of two weeks.

Mini-project ideas : 

  • detect redundancy in lyrics of songs over the years
  • Music visualization (e.g. :
  • Scrape, mine and consolidate a database for Hindi Poetry, bandishes and Carnatic songs