zhixin piao - piaozhx.com filethe 1st prize in china undergraduate mathematical contest in modeling...

2
ZHIXIN PIAO Computer Vision, Machine Learning, Deep Learning (+86) 17621504831 [email protected] www.github.com/piaozhx www.piaozhx.com EDUCATION ShanghaiTech University Shanghai China School of Information Science and Technology, M.S. in Computer Science Sep. 2017 - Present Advisor: Prof. Shenghua Gao, Major in Computer Vision SouthEast University Nanjing China School of Computer Science and Engineering, B.S. in Computer Science Sep. 2013 - Jun. 2017 Advisor: Prof. Guilin Qi, Major in Data Mining SKILL Research Insterent: Image Synthesis, Human Pose Estimation, Trajectory Prediction, Image Sementation Programming: Python(Pytorch), C++, Matlab, JS, CSS, HTML Knowledge: Tornado, Bootstrap, Docker, Git WORK EXPERIENCE Tencent Youtu Lab Shanghai China Computer Vision Research Intern Nov. 2018 - Match. 2019 ShanghaiTech University School of Information Science and Technology Shanghai China CS172 - Computer Vision I (Fall 2018) Teaching Assistant Sep. 2018 - Jan. 2019 ShanghaiTech University School of Information Science and Technology Shanghai China High Performance Cluster (HPC) DevOps Assistant May. 2018 - Present PUBLICATIONS (* INDICATES EQUAL CONTRIBUTION) Motion Imitation + Source Image Reference Pose Synthesized Image Appearance Transfer + Source Image Reference Appearance Synthesized Image Novel View Synthesis + Source Image Novel Camera Synthesized Image Liquid Warping GAN: A Unified Framework for Human Motion Imitation, Appearance Transfer and Novel View Synthesis [Github] [Project Page] Wen Liu * , Zhixin Piao * , Jie Min, Wenhan Luo, Lin Ma, Shenghua Gao ICCV 2019 Introduce a 3D body parametric model to disentangle pose and shape which provides more informa- tion with details than 2D pose Propose a unified framework for human motion imitation, appearance transfer and novel view syn- thesis and design a Liquid Warping Block to preserve the source identity and address the loss of source information Build a new dataset for the evaluations on human motion imitation, appearance transfer and novel view synthesis Encoding Crowd Interaction with Deep Neural Network for Pedestrian Trajectory Pre- diction [Github] Yanyu Xu * , Zhixin Piao * , Shenghua Gao CVPR 2018 Propose CIDNN to extractor spatial and temporal feature from multiple object attention Best performance on multiple popular dataset (GC, Subway by CUHK etc.) Easy to re-implement and fast(1.91 ms/f on CPU, 0.43 ms/f on GPU) Parsing-specific Feature Extractor Common Feature Extractor Pose-specific Feature Extractor 3D Human Body Modulation Module Feature Consolidation Module Feature Consolidation Module Parsing Task 2D Pose Estimation Task 3D Human Body Element-wise Add operation Conv 3x3 Conv Offsets Offset field 2N Deformable Convolution (c) CD-Conv Module par f pos f Feature Extractor Feature Extractor CD-Conv CD-Conv (a) The Overall Network Architecture Stage-I Feature Separation Stage Stage-II Feature Union Stage HMR Rendering Mask Heatmap f f union par L union pos L (b) Feature Consolidation Module coarse par L fine par L par f com f Concat par P d Deconv Deconv coarse par P + fine coarse par par par P P P d = fine pos L fine par L 1x1 Conv concat Avg. Pool FC Gumbel Samples Softmax Argmax Forward Backward Image Feature Extractor Image Feature Extractor 1x1 Conv Common feature Gate Gate (e) 3D Human Body Modulation Module If argmax is 1, common feature adds SMPL heatmap and mask feature SUNNet: A Novel Framework for Simultaneous Human Parsing and Pose Estimation Yanyu Xu, Zhixin Piao, Shenghua Gao Neurocomputing (under review) Propose SUNNet, which encodes the correlation between parsing and pose estimation both implicitly and explicitly Leverage 3D human body reconstructed from a single image to enhance the performance of human parsing and pose estimation Extensive experiments validate the effectiveness of our method for joint human parsing and pose estimation on the LIP dataset 1

Upload: others

Post on 21-Sep-2019

8 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: ZHIXIN PIAO - piaozhx.com fileThe 1st Prize in China Undergraduate Mathematical Contest in Modeling (CUMCM), Jiangsu Province Jul. 2015 The 2nd Prize(Honorable Mention) in Mathematical

ZHIXIN PIAOComputer Vision, Machine Learning, Deep Learning

(+86) 17621504831 • [email protected] • www.github.com/piaozhx • www.piaozhx.com

EDUCATION

ShanghaiTech University Shanghai ChinaSchool of Information Science and Technology, M.S. in Computer Science Sep. 2017 - PresentAdvisor: Prof. Shenghua Gao, Major in Computer Vision

SouthEast University Nanjing ChinaSchool of Computer Science and Engineering, B.S. in Computer Science Sep. 2013 - Jun. 2017Advisor: Prof. Guilin Qi, Major in Data Mining

SKILL

Research Insterent: Image Synthesis, Human Pose Estimation, Trajectory Prediction, Image Sementation

Programming: Python(Pytorch), C++, Matlab, JS, CSS, HTML

Knowledge: Tornado, Bootstrap, Docker, Git

WORK EXPERIENCE

Tencent Youtu Lab Shanghai ChinaComputer Vision Research Intern Nov. 2018 - Match. 2019

ShanghaiTech University School of Information Science and Technology Shanghai ChinaCS172 - Computer Vision I (Fall 2018) Teaching Assistant Sep. 2018 - Jan. 2019

ShanghaiTech University School of Information Science and Technology Shanghai ChinaHigh Performance Cluster (HPC) DevOps Assistant May. 2018 - Present

PUBLICATIONS (* INDICATES EQUAL CONTRIBUTION)

Motion Imitation

+

Source Image Reference Pose Synthesized Image

Appearance Transfer

+

Source Image Reference Appearance Synthesized Image

Novel View Synthesis

+

Source Image Novel Camera Synthesized Image

Liquid Warping GAN: A Unified Framework for Human Motion Imitation, AppearanceTransfer and Novel View Synthesis [Github] [Project Page]Wen Liu∗, Zhixin Piao∗, Jie Min, Wenhan Luo, Lin Ma, Shenghua Gao ICCV 2019

• Introduce a 3D body parametric model to disentangle pose and shape which provides more informa-tion with details than 2D pose

• Propose a unified framework for human motion imitation, appearance transfer and novel view syn-thesis and design a Liquid Warping Block to preserve the source identity and address the loss ofsource information

• Build a new dataset for the evaluations on human motion imitation, appearance transfer and novelview synthesis

Encoding Crowd Interaction with Deep Neural Network for Pedestrian Trajectory Pre-diction [Github]Yanyu Xu∗, Zhixin Piao∗, Shenghua Gao CVPR 2018

• Propose CIDNN to extractor spatial and temporal feature from multiple object attention

• Best performance on multiple popular dataset (GC, Subway by CUHK etc.)

• Easy to re-implement and fast(1.91 ms/f on CPU, 0.43 ms/f on GPU)

Parsing-specific

Feature

Extractor

Common

Feature

Extractor

Pose-specific

Feature

Extractor

3D Human Body

Modulation

Module

Feature

Consolidation

Module

Feature

Consolidation

Module

Parsing Task2D Pose

Estimation Task3D Human Body Element-wise

Add operation

Conv

3x3

Conv

Offsets

Offset field

2N

Deformable Convolution

(c) CD-Conv Module

parf

posf

Feature

Extractor

Feature

ExtractorCD-Conv

CD-Conv

(a) The Overall Network Architecture

Stage-I Feature Separation Stage Stage-II Feature Union Stage

HMR

Rendering

Mask Heatmap

parf

posf

union

parL

union

posL

(b) Feature Consolidation Module

coarse

parLfine

parL

parf

comf

Concat

parPdDeconv

Deconvcoarse

parP +fine coarse

par par parP P Pd=

fine

posL

fine

parL

1x1

Conv

concat

Avg.

Pool FC

Gumbel

SamplesSoftmax

ArgmaxForward

Backward

Image

Feature

Extractor

Image

Feature

Extractor

1x1

Conv

Common

feature

Gate

Gate

(e) 3D Human Body Modulation Module

If argmax is 1, common feature adds

SMPL heatmap and mask feature

SUNNet: A Novel Framework for Simultaneous Human Parsing and Pose EstimationYanyu Xu, Zhixin Piao, Shenghua Gao Neurocomputing (under review)

• Propose SUNNet, which encodes the correlation between parsing and pose estimation both implicitlyand explicitly

• Leverage 3D human body reconstructed from a single image to enhance the performance of humanparsing and pose estimation

• Extensive experiments validate the effectiveness of our method for joint human parsing and poseestimation on the LIP dataset

1

Page 2: ZHIXIN PIAO - piaozhx.com fileThe 1st Prize in China Undergraduate Mathematical Contest in Modeling (CUMCM), Jiangsu Province Jul. 2015 The 2nd Prize(Honorable Mention) in Mathematical

Cascaded ConvLSTMs using Semantically-Coherent Data Synthesis for UnsupervisedVideo Object SegmentationJia Zheng, Weixin Luo, Zhixin Piao IEEE Access

• Propose Stacked-ConvLSTM and Cascade module for unsupervised Video Object Segmentation

• First RGB based feature(without optical flow) work on this task

• a new data augmentation to overcome small dataset problem

Entity Linking in Web Tables with Multiple Linked Knowledge BasesTianxing Wu, Shengjia Yan, Zhixin Piao, Liang Xu, Ruiming Wang, Guilin Qi JIST 2016

• Propose a random-walking based algorithm for Entity Linking in web tables

PROJECT

Context Awared Object Tracking By Deep Reinforcement Learning Shnaghai ChinaCourse Project, ShanghaiTech University(CS280 Deep Learning) Dec. 2017

• Implement Correlation-Filter algorithm by multiple feature(HOG, Histogram)

• Propose a Context Awared object tracking method by Deep Reinforcement Learning(A3C)

Docker Monitor and Manager System for Deep Learning Cluster Shanghai ChinaDevOps Project [Github] Sep. 2018

• Build a deep learning developing environment (Including Tensorflow, Pytorch, Mxnet...)

• Build a container manager system(based on tornado, mariaDB, bootstrap, mkDocs...)

• Exclude it to multiple user container system

AWARDS AND HONORS

The 1st Prize in China Undergraduate Mathematical Contest in Modeling (CUMCM), Jiangsu Province Jul. 2015

The 2nd Prize(Honorable Mention) in Mathematical Contest in Modeling (MCM), America Feb. 2016

The 3rd Prize in Collegiate Programming Contest, Jiangsu Province May. 2016

2