Site Network: Personal | My Company | Vision Research Lab @ UCSB |

My Research

I first became acquainted with the domain of image analysis during my last year in college. Since then I have been exploring some of the relevant problems in this field.

I like principled approaches that start from a solid theoretical framework and that finally become working implementations.

Alas, I do not always succeed... but at least I have fun also when I am working!


Corner Detection

with Dr. C. Kenney and Prof. B. S. Manjunath

(2003-now) In this project we developed a unified framework to analyze different corner detectors that are based on the spectral properties of the image autocorrelation matrix. The main mathematical tool used to carry out our analysis is based on the condition theory. Using this approach we were able to show that some of the most commonly used corner detectors can be unified by choosing an appropriate matrix norm. The mathematical framework we developed allowed us to easily extend some of the commonly used detectors to images with pixel dimension greater than 2 (i.e. tomographic images) and intensity dimension greater than one (like RGB or multispectral images). We are now using these tools to identify which relevant properties are shared by some of the commonly used corner detectors and which ones make them unique. Moreover we are interested in applying the condition theory framework to detect the intrinsic structure of a neighborhood of an image point. This is leading us to develop a multiscale condition theory based point detector framework.


Curves & Curve Descriptors

with Dr. C. Kenney, L. Bertelli, Prof. S. Chandrasekaran and Prof. B. S. Manjunath

(2003-2005) In this project we are investigating efficient methods to detect, describe and match curves, despite the geometric distortions that arise when the scene is imaged from different points of view. In this context we developed a normalization procedure that allows to extract the shape of a curve independently from an affine transformations of the curve itself. We also introduced a physically motivated descriptor based on the Helmholtz equation to label the curves for matching and retrieval purposes. Current research is focused on alleviating the computational burden for the detection and description of the curves and in combining together shape and content of image patches. We are also intrested in extending our methods for 3D surfaces.



with Dr. C. Kenney, Dr. M. Bober and Prof. B. S. Manjunath

(2004-now) In this project we are trying to construct a RANSAC framework that will enable us to perform parameter estimation robustly in different scenarios characterized by the presence of large quantities of outliers. We are developing methods that will speed up the convergence of the traditional algorithm, that will allow us to perform the fusion of information coming from different sources and that can cope with the presence of multiple models. We are also interested in characterizing the stability of the solutions found by RANSAC.


Image Equalization, Blending & Stitching

with Prof. B. S. Manjunath

(now) In this project we propose methods to produce seamless mosaics from registered images. We investigate three steps. The first one is the robust equalization of the images (using the RANSAC framework). The second step is the identification of an appropriate stitching curve using a physically motivated method based on the propagation of waves in a non uniform medium. Finally the last step aims at blending the images in the neighborhood of the seam using a wavelet decomposition.


Multiview Curve Database

with Dr. C. Kenney, Dr. S. Bhagavathy and Prof. B. S. Manjunath

(2004) The Multiview Curve Database (MCD) has been constructed  to test the performance of  shape/contour descriptors in the presence of perspective distortions. The images of 40 shapes (extracted from the MPEG-7 shape database) have been taken under seven different points of view. The contour of each object has been extracted, an arbitrary rotation has been applied and 7 new contours have been generated by mirroring the original ones. The database consists of 14 contours for each of the 40 objects.


Head Tracking

with Prof. R. Frezza and Prof. B. S. Manjunath

(2000-2001) In this project we implemented an head tracking system able to recover the trajectory of a human head in the space (i.e. for each frame the position and orientation of the head). The algorithm uses as only input the video stream obtained by one single camera. The algorithm is markerless, it is able to track also fairly complex movements of the head, it is quite robust under varying light conditions and it is easy to initialize.