MS Thesis

Learning semantic interaction among indoor objects


Support order prediction in clutter for manipulation
In this work, support relationship among objects involving physical contact is inferred and a sequence or support order is learned in which the surrounding objects of our object of interest should be removed while causing minimal damage to the environment. Support order prediction is also extended to multiple views to overcome the limitations of single-view. An RGBD dataset using Kinect is collected to explore clutter involving physical contact. Details about the project and the link to RGBD dataset for download can be found here.


Selected Projects


Semantic Segmentation: Learn from Neighbourhood
[Course Project | Oct’15-Dec’15]
Performed segmentation on RGBD data using signature properties of a region such as uniform surface, depth gradient, or color, CRFs and belief propagation.

3D Shape Completion for Graspable Objects
[Ongoing Research Project at NEU | Sep’15-Current]
Currently working on shape completion of novel objects using exemplar training method by Rock et. al. The aim of this project is to recover or rectify missing information about an object, given a noisy point cloud.

Object Detection on RGBD Images
[Research Project at NEU | Jan-Aug’15]
Implemented object detection on the "University of Washington RGBD Dataset", and compared the performance with the state-of-the-art performance. I explored different features, machine learning techniques and techniques to generate object proposals. In my final implementation, I used Selective Search, Deformable Parts Model, Multi-class SVM and Hard Negative Mining to obtain the best performance.

Point Cloud Scene Registration
[Research Project at NEU | Oct-Dec’14]
Explored different descriptors and approaches to solve the problem of ``scene registration'' in RGBD scenario, where holes and noisy data pose significant challenge.

Object Localization in Clutter for Grasping
[Research Project at NEU | Sep-Oct’14]
Performed localization and pose estimation of an object in clutter given the model of the object. This was followed by localizing grasp affordances on the object and subsequently grasping the object by the Baxter robot.

Obstacle Detection on Floor using Kinect
[Independent Project | Aug-Sep’13]
Performed floor plane segmentation followed by segmentation of objects lying on the plane. The dimensions of the objects were calculated and the objects were displayed with colors according to the distance from camera.

Movie Scene Localization
[Research Project at CVIT | Oct-Dec’11]
This project aimed at retrieving a location appearing in various scenes of one or more movies using visual Bag-of-words approach.

Image Retrieval
[Course Project | Feb-Mar’11 | Team Size: 2]
Developed a complete pipeline for image retrieval using Bag of Words model on Caltech-101 image dataset and further improved its performance using tf-idf and query expansion.

Video Analysis
[Course Project | Oct-Dec’10 | Team Size: 2]
Analyzed long news video footage using computer vision techniques such as shot detection, scene classification and face detection, and performed tasks such as news channel detection, news tickers removal, news anchor detection, cricket shot detection, parliament news classification, etc.

Bidirectional Similarity
[Course Project | Oct-Dec’10 | Team Size: 2]
Implemented the paper titled "Summarizing Visual data using bidirectional similarity" by Simakov et al. in a group of two. In this work, image re-targeting was performed using bidirectional similarity that preserved completeness and coherence.

Wavelet-based Image Denoising
[Course Project | Oct-Dec’10 | Team Size: 2]
Implemented the work "Adaptive wavelet thresholding for image denoising and compression" by Chang et al. In this work, Visushrink (universal thresholding) and BayesShrink (adaptive wavelet-based soft thresholding) are applied to noisy images and compared.

Real-time Tracking for Surveillance
[Research Project at IIT Madras | Jan-Jun’10 | Team Size: 2]
Real-time objects/humans tracking in outdoor setting was performed using Kalman filter as part of the project “Multiple Camera Detection and Tracking”.

Color Image Segmentation
[Independent Project | May-Jun’09]
Implemented Color Image Segmentation using "Unsupervised approach" and "Euclidean Distance Method" and performed a comparison of the two approaches based on the paper titled "Unsupervised approach to color video thresholding", by Du \textit{et al}.

Networking Protocols
[Development Project at Aricent | Feb’08-Jan’10 ]
Worked in multiple projects involving development of multicasting protocols in layer 3 such as PIM, IGMP and DVMRP and stabilization of automated test suites for layer 2 protocols.