My current research interest lies in solving theoretical and practical problems entailing Computer Vision and Machine Learning, specifically on the problems related to image co-segmentation, remote sensing, medical imaging and affective computing. Our group works extensively in Deep Learning, not only exploring innovative application of deep learning on existing problems but specially extending its theoretical understanding.
Have a look at http://im.itu.edu.pk/research for our research and projects in the areas of Deep Learning, Remote Sensing and Affective Computing.
On the journey called PhD, I dipped my toes in many different problems. In the last year or so me and adviser have introduced concept Decsontructive Learning. One of the problem that kept intriguing me was discovering an object automatically from the set of images. We explored the problem with respect to the Multi-image Co-segmentation. I explored Sparse represnetation and dictionary learning to specifically see how it could be used for image classification. We introduced part wise affine constraints that could handle images with noise and/or occlusion. Selected publications are listed here.
We introduced the novel notion of deconstructive learning and proposed a practical computational framework for deconstructing a broad class of binary classifiers commonly used in computer vision applications. While the ultimate objective of most learning problems is the determination of classifiers from labeled training data, for deconstructive learning, the objects of study are the classifiers themselves. As its name suggests, the goal of deconstructive learning is to deconstruct a given classifier by determining and characterizing (as much as possible) the full extent of its capability, revealing all of its powers, subtleties and limitations. In particular, this problem is motivated by the seemingly innocuous question that given an image-based binary classifier C as a black-box oracle, how much can we learn of its internal working by simply querying it?
I am working on the problem of Multi-image Co-segmentation, (segmenting out object similar in the given set of images). Two important words here are “object” and the “similar”. Previous works have looked into the problem by segmenting out similar region and have defined “similar” to be similarity in color. This similarity is represented as energy function and problem is formulated as engergy minimization problem.
Such a structure does not allow more control over what we want to be segment out. We, on the other hand, are iterested in segmenting out “object” that might have similar color, shape or structure. For this we segment out each image at different speed, easy images get segmented out much faster than complex images (images with many objects and texture) and then after each iteration share information about the images. Thus helping segment out complex images from information we have gathered from the easy images.
Dictionary Learning and Sparse Representation
Worked on the problem of dictionary learning w.r.t sparse representation of data. We explored use of Block and Group structures and introduced affine constraints for image classification. This work resulted in publications listed below.