Aroma Mahendru

              

Research Interests

I am interested in intersection of machine learning and computer vision especially object detection and graphical models.

Major Projects

Object-Proposal Evaluation Protocol is 'Gameable'

In this project, we demonstrate that the evaluation protocol for the evaluation of object proposals is biased. We conduct a thorough evaluation of various proposal methods on three datasets. We also conduct experiments to introduce diagnostic tools to alleviate this bias.

Robust Reading in Scene Images

A classifier and a proposal generator specific to text in urban scene images were developed as a part of this project. [Report]

Bethe ADMM and modeling budget factors for ADMM

As part of graphical models toolkit, Bethe-ADMM was implemented for Graphlab. A special type of factor, called budget factor was introduced as a feature to ADMM graph. [Code]

Implementation of Dual Decomposition and AD3 on Graphlab

Symmetric and Projected Dual Decomposition, Alternating Direction Dual Decomposition were implemented on Graphlab. [Code]

Classification of Transparent objects via Polarization Imaging

Algorithm was developed to classify transparent objects from their surroundings using a special property of specular reflection. [Publication]

Relevant Links

Github page