Stacked dense-hourglass networks for human pose estimation

Wang, Dongbo

Permalink

https://hdl.handle.net/2142/101155 Copy

Description

Title

Stacked dense-hourglass networks for human pose estimation

Author(s)

Wang, Dongbo

Issue Date

2018-04-12

Director of Research (if dissertation) or Advisor (if thesis)

Schwing, Alexander

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Date of Ingest

2018-09-04T20:34:02Z

Keyword(s)

• Stacked Hourglass Networks
• DenseNets
• Pose Estimation

Abstract

Convolutional Neural Networks (CNNs) are driving major advances in many computer vision tasks, including the problem of 2D single-person pose estimation. For this task, the Stacked Hourglass Networks (Stack-HgNets) is one of the state-of-the-art architecture that uses residual modules extensively as the basic building block. The residual modules are well recognized for creating shortcut connections, skipping one or more layers which allows information and gradients to flow more effectively through a deep network without vanishing. In this work, we build on the Stack-HgNets and introduce the Stacked Dense-Hourglass Networks (Stack-DenseHgNets). They use dense blocks instead of the residual modules as the basic building block. The dense blocks create more direct connections between each layer and its subsequent successors, granting later filters the access to all the preceding feature-maps inside the same block. Therefore, dense blocks serve as the upgraded substitution for the residual modules. We evaluate the Stack-DenseHgNets on the popular human pose estimation benchmark dataset and compare its performance to the original Stack-HgNets. Using fewer parameters, the Stack-DenseHgNets obtains a performance competitive to the state-of-the-art results on the MPII Human Pose Dataset.

Graduation Semester

2018-05

Type of Resource

text

Permalink

http://hdl.handle.net/2142/101155

Copyright and License Information

Copyright 2018 Dongbo Wang

Owning Collections