I am a PhD student at the University of Maryland, College Park in the department of Computer Science. I am currently working with the Computer Vision Lab associated with the Institute for Advanced Computer Studies (UMIACS), with Prof. Larry Davis as my advisor. I completed my BTech in Electrical Engineering at the Indian Institute of Technology, Kanpur. My research interests are primarily in Computer Vision and Machine Learning.
I am currently working on outlier detection and multi-classification for datasets containing a fairly large number of classes (~50). The major focus here is on datasets have a very high feature dimensionality (~15000), but a fairly low number of samples per class (~100). This makes the task challenging, since it becomes very hard to identify (even a criteria for detecting) outliers due to the induced multicollinearity. The key focus on solving this problem, is in the use of our recently developed "Composite Discriminant Factor Analysis," (which is a variant of Partial Least Squares) to project the data on to a substantially low dimensional subspace, which preserves even more discriminating information than conventional PLS. More details about the project can be found here.
I worked on tracking stationary vehicles in wide angle satellite videos. More details about the project can be found here.
The full body detector based on HOG (Histogram of Oriented Gradients) features is trained on the INRIA pedestrian dataset using a novel technique of "Latent Partial Least Squares(LPLS)". The technique is an adaptation of the standard NIPALS PLS-1 Algorithm to the acclaimed Latent Support Vector Machine (LSVM) algorithm which won the 2010 PASCAL "Lifetime Achievement" prize for object detection.
The word "Latent" in both LSVM and LPLS refers to the fact that the object bounding box annotations in the training set are not perfect. For example, a bounding box for a person might not be tight or alternatively a bounding box may be such that a person is not exactly at its center. For object recognition techniques, when one uses features which are highly location specific [ex: HOG features], it is very important for the training bounding boxes from various images to be mutually registered. However, standard registration techniques can't be used here since each bounding box corresponds to a different person having different appearances. Latent-PLS mitigates this problem by making the bounding box locations "latent" or hidden variables and applying the process of alternating between choosing optimal values for these variables and updating the model.
Vlad Morariu, Ejaz Ahmed, Venkataraman Santhanam, David Harwood and Larry Davis. Composite Discriminant Factor Analysis. 2014 IEEE Winter Conference on Applications of Computer Vision (WACV)[PDF]
Arvind Tolambiya, Venkataraman Santhanam, and Prem K. Kalra. Content-based image classification with wavelet relevance vector machines. SOFT COMPUTING - A FUSION OF FOUNDATIONS, METHODOLOGIES AND APPLICATIONS Volume 14, Number 2 (2010), 129-136. [PDF]
Department of Computer Science
University of Maryland, College Park