A 'good' travel model could yield new scientific insights into human behavior, aside for being useful in numerous applications. For example, travel models can help in predicting the spread of disease; surveying tourism, traffic, and special events mobility for urban planning; geolocating with computer vision; interpreting activity from movements; and recommending travel. The objective of our project was to implement a basic latent travel model. Essentially, we model travel probabilities as functions of spatially-varying latent properties of locations and travel distance. The latent factors represent interpretable properties: travel distance cost and desirability of locations. These latent factors are combined in a multiplicative model which easily lends itself to incorporating additional latent factors and sources of information.

Estimating geographic information from an image is an excellent, difficult high-level computer vision problem with applicability in a wide array of disciplines. Intuitively, one can imagine that the socioeconomic status of an individual would play a key role in determining the individual's travel propensity. However, due to the unavailability of such a dataset, determining the socioeconomic status of an individual is a non-trivial task. For example, most recent works in this area have used the publicly-available Flickr.com image streams of individuals to build their travel models. In other words, our task is reduced to learning the socioeconomic status of an individual from his/her uploaded photographs. One factor that may be used to determine the socioeconomic status of an individual is the type of restaurants he or she frequents. For example, an individual who regularly frequents expensive restaurants is likely to be in a higher socioeconomic bracket than one who eats at fast food joints.

The goal of this project was to implement and thorougly analyze the image segmentation algorithm proposed by Felzenszwalkb et. al. (2004). The proposed method defines a metric for measuring the evidence of a boundary between two regions using a graph-based representation of the image. Based on the proposed metric, an efficient image segmentation algorithm is developed. Although this algorithm is a greedy algorithm, it respects some global properties of the image. Some important features of the proposed algorithm are that it runs in linear time and that it has the capability to adapt its behavior differently between regions of high-variability and low-variability. In particular, we demonstrate that it ignores details in high-variability regions.

In 2012, I spent a wonderful summer interning at the Technischen Universität Berlin, with the Neural Information Processing Group under the guidance of Dr. Johannes Mohr and Prof. Klaus Obermayer. I worked on developing a recognition system that is largely scale-, illumination-, orientation-invariant and can be used on any object regardless of its shape or size. The recognition system is also capable of recognizing objects in a cluttered scene in almost real time.

In 2011, I interned at the Defence Research & Development Organization, India with the Defence Terrain Research Lab. During my time there I developed CTD, a unique feature descriptor,for color image retrievel and browsing. A key goal of the project was to allow satellite image browsing which can be used to detect military installations.

Our goal was to study and analyze the patterns observed during IP packet transfers through various locations around the world. Packets take different routes to reach a specific destination and our goal was to observe the variation in their latencies and check if there is any behavioral pattern for traces on weekdays and weekends. Also we studied the dependence of latency with respect to distance of the destination from source and for different countries. Additionally, we observed that there are at times transatlantic hops in some ‘local’ pings and we reported the observation and its impact it has over the network.

During my final year of undergraduate studies, I worked on American Sign Language recognition using the Microsoft Kinect camera. I also designed and implemented specific GUI applications, such as paint, using hand gestures as input.

The following application can be used to generate a random roadmap for a given scene. Notes on installation/compilation instructions for win 32 and further details on usage can be found in the Readme.