Ph.D. Dissertation Defense: Raviteja Vemulapalli

Monday, August 8, 2016
3:00 p.m.
Room 3450 AVW Bldg.
Maria Hoo
301 405 3681
mch@umd.edu

ANNOUNCEMENT: Ph.D. Dissertation Defense

Name: Raviteja Vemulapalli
 
Advisory Committee:
Professor Rama Chellappa, Chair/Advisor
Professor Larry S. Davis
Professor Min Wu
Professor Amitabh Varshney
Professor Ramani Duraiswami
 
Date/Time:  Monday, August 8th, 2016 at 3pm

Location: Room 3450 A. V. Williams building

Title: Geometric Representations and Deep Gaussian Conditional Random Field Networks for Computer Vision

Representation and context modeling are two of the most important factors that affect the performance of computer vision algorithms. In this dissertation, we focus on the representation and context modeling issues for some computer vision problems and make novel contributions by proposing new 3D geometry-based feature representations for recognizing human actions from skeletal sequences, and introducing Gaussian conditional random field model-based deep network architectures that explicitly model the spatial context by considering the interactions among the output variables.
 
data-based human action recognition. While numerous approaches have been proposed in the past for this task, most of the existing works represent a human skeleton by using either the joint coordinates or the joint angles, and focus on modeling the temporal evolution of the skeletal representation and classifying it. In contrast to these works, here, we try to answer the following fundamental question: How should we represent a 3D human skeleton for action recognition? Inspired by the observation that for human actions, the relative geometry between various body parts provides a more meaningful description than their absolute locations, we propose new skeletal representations that explicitly model the relative 3D geometry between all pairs of body parts. The proposed geometric representations result in skeletal features that are members of the special orthogonal and special Euclidean groups, and action representations that are curves in these Lie groups. Since classification of action curves in these non-Euclidean spaces is a difficult task, we map the action curves to the corresponding Lie algebras, which are vector spaces, using appropriate techniques and perform action recognition by classifying the Lie algebra curves. Experimental results on various human action recognition datasets show that the proposed representations clearly outperform various existing skeletal representations in terms of action recognition accuracy.
 
In the second part of this thesis, we focus on the aspect of context modeling and propose novel deep networks that explicitly model the interactions between output variables. In the past few years, deep networks have revolutionized the field of computer vision improving the state-of-the-art results in various applications by a huge margin. However, standard feed-forward networks do not explicitly model the interactions between output variables, which is important in various applications such as semantic image segmentation, object detection, etc. Traditionally, graphical models, especially Conditional Random Field (CRF) models, have been widely-used to model the interactions between output variables. In this work, we combine both these ideas and propose feed-forward deep networks based on Gaussian CRF models. The proposed network architecture consists of two sub-networks, a parameter generation network (PgNet) that generates the parameters of the Gaussian CRF model, and an inference network (InfNet) that performs approximate inference of the Gaussian CRF model for the parameters generated by the PgNet. We propose two Gaussian CRF-based deep networks under this architecture, one for semantic image segmentation and the other for image denoising, and both these networks achieve state-of-the-art results when trained end-to-end using gradient-based optimization techniques.
 

Audience: Graduate  Faculty 

remind we with google calendar

 

March 2024

SU MO TU WE TH FR SA
25 26 27 28 29 1 2
3 4 5 6 7 8 9
10 11 12 13 14 15 16
17 18 19 20 21 22 23
24 25 26 27 28 29 30
31 1 2 3 4 5 6
Submit an Event