東華大學圖書館 |

語系: 繁體中文

說明(常見問題)

回圖書館首頁

手機版館藏查詢

登入

回首頁

切換: 標籤 | MARC模式 | ISBD

Articulated Human Pose Estimation wi...

Yang, Wei.

FindBook

Google Book

Amazon

博客來

Articulated Human Pose Estimation with Deep Learning.

紀錄類型:	書目-電子資源 : Monograph/item
正題名/作者:	Articulated Human Pose Estimation with Deep Learning./
作者:	Yang, Wei.
出版者:	Ann Arbor : ProQuest Dissertations & Theses, : 2018,
面頁冊數:	122 p.
附註:	Source: Dissertations Abstracts International, Volume: 80-06, Section: B.
Contained By:	Dissertations Abstracts International80-06B.
標題:	Computer Engineering. -
電子資源:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=11012264
ISBN:	9780438659667

Articulated Human Pose Estimation with Deep Learning.
Yang, Wei.

Articulated Human Pose Estimation with Deep Learning. - Ann Arbor : ProQuest Dissertations & Theses, 2018 - 122 p.

Source: Dissertations Abstracts International, Volume: 80-06, Section: B.

Thesis (Ph.D.)--The Chinese University of Hong Kong (Hong Kong), 2018.

This item must not be sold to any third party vendors.

Articulated human pose estimation is a fundamental yet challenging task in computer vision. The goal is to estimate 2D or 3D locations of human body joints given an image or a video sequence. It serves as informative knowledge for higher-level applications, such as activity recognition, human-computer interaction, robotics vision, and autonomous driving. Although promising progress has been achieved by deep neural networks, obtaining accurate human pose estimation is still nontrivial due to the highly articulated human body limbs, occlusion, cluttered background, scale variation, and foreshortening. To cope with these challenges, two key ingredients, i.e., the structures of the articulated human body, and the feature representation, should be carefully considered. In this thesis, we propose to tackle human pose estimation with deep learning method enhanced by incorporating these three key ingredients. To model the structure of the human body, we first propose an end-to-end trainable framework which incorporates domain prior knowledge, i.e., the geometric relationships among human body joints, into the powerful deep convolutional neural networks. This structural modeling greatly regularizes the training process and is especially effective on benchmarks with challenging articulations. On the other hand, to learn a better feature representation which is robust to scale variation, occlusion, and cluttered background, we further propose to learn feature pyramid within the traditional deep neural networks. Specifically, we design a pyramid residual module. It is a multi-branch architecture and learns convolutional filters on various scales. Additionally, we observe that traditional weight initialization methods for networks with multiple branches are inappropriate. Therefore, we provide a theoretically sound extension of the current weight initialization schemes for multi-branch networks. Experiments demonstrate the effectiveness of our methods on various tasks include human pose estimation and image classification. Although significant advances have been achieved in 2D pose estimation due to the availability of large-scale datasets with images in the wild, the progress in 3D human pose estimation remains limited. Existing datasets are mostly collected by motion capture sensors in the constrained lab environment, hence the learned models have very poor generalization ability on images in the wild. In this thesis, we propose an adversarial learning framework, which distills the 3D human pose structures learned from the fully annotated dataset to in-the-wild images with only 2D pose annotations. Instead of defining hard-coded rules to constrain the pose estimation results, we design a novel multi-source discriminator to distinguish the predicted 3D poses from the ground-truth, which helps to enforce the pose estimator to generate anthropometrically valid poses even with images in the wild.

ISBN: 9780438659667Subjects--Topical Terms:

1567821
Computer Engineering.
Subjects--Index Terms:

Adversarial learning

Articulated Human Pose Estimation with Deep Learning.
LDR:04316nmm a2200409 4500 001 2272027
005 20201102111649.5
008 220629s2018 ||||||||||||||||| ||eng d
020 $a 9780438659667
035 $a (MiAaPQ)AAI11012264
035 $a AAI11012264
040 $a MiAaPQ $c MiAaPQ
100 1 $a Yang, Wei. $3 1257167
245 1 0 $a Articulated Human Pose Estimation with Deep Learning.
260 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2018
300 $a 122 p.
500 $a Source: Dissertations Abstracts International, Volume: 80-06, Section: B.
500 $a Publisher info.: Dissertation/Thesis.
502 $a Thesis (Ph.D.)--The Chinese University of Hong Kong (Hong Kong), 2018.
506 $a This item must not be sold to any third party vendors.
506 $a This item must not be added to any third party search indexes.
506 $a This item is not available from ProQuest Dissertations & Theses.
520 $a Articulated human pose estimation is a fundamental yet challenging task in computer vision. The goal is to estimate 2D or 3D locations of human body joints given an image or a video sequence. It serves as informative knowledge for higher-level applications, such as activity recognition, human-computer interaction, robotics vision, and autonomous driving. Although promising progress has been achieved by deep neural networks, obtaining accurate human pose estimation is still nontrivial due to the highly articulated human body limbs, occlusion, cluttered background, scale variation, and foreshortening. To cope with these challenges, two key ingredients, i.e., the structures of the articulated human body, and the feature representation, should be carefully considered. In this thesis, we propose to tackle human pose estimation with deep learning method enhanced by incorporating these three key ingredients. To model the structure of the human body, we first propose an end-to-end trainable framework which incorporates domain prior knowledge, i.e., the geometric relationships among human body joints, into the powerful deep convolutional neural networks. This structural modeling greatly regularizes the training process and is especially effective on benchmarks with challenging articulations. On the other hand, to learn a better feature representation which is robust to scale variation, occlusion, and cluttered background, we further propose to learn feature pyramid within the traditional deep neural networks. Specifically, we design a pyramid residual module. It is a multi-branch architecture and learns convolutional filters on various scales. Additionally, we observe that traditional weight initialization methods for networks with multiple branches are inappropriate. Therefore, we provide a theoretically sound extension of the current weight initialization schemes for multi-branch networks. Experiments demonstrate the effectiveness of our methods on various tasks include human pose estimation and image classification. Although significant advances have been achieved in 2D pose estimation due to the availability of large-scale datasets with images in the wild, the progress in 3D human pose estimation remains limited. Existing datasets are mostly collected by motion capture sensors in the constrained lab environment, hence the learned models have very poor generalization ability on images in the wild. In this thesis, we propose an adversarial learning framework, which distills the 3D human pose structures learned from the fully annotated dataset to in-the-wild images with only 2D pose annotations. Instead of defining hard-coded rules to constrain the pose estimation results, we design a novel multi-source discriminator to distinguish the predicted 3D poses from the ground-truth, which helps to enforce the pose estimator to generate anthropometrically valid poses even with images in the wild.
590 $a School code: 1307.
650 4 $a Computer Engineering. $3 1567821
650 4 $a Artificial intelligence. $3 516317
650 4 $a Computer science. $3 523869
653 $a Adversarial learning
653 $a Computer vision
653 $a Deep learning
653 $a Feature learning
653 $a Graphical model
653 $a Human pose estimation
690 $a 0464
690 $a 0800
690 $a 0984
710 2 $a The Chinese University of Hong Kong (Hong Kong). $b Electronic Engineering. $3 2094192
773 0 $t Dissertations Abstracts International $g 80-06B.
790 $a 1307
791 $a Ph.D.
792 $a 2018
793 $a English
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=11012264