10
Bing Yu (COMP8755) Supervisor: Jo Plested, Tom Gedeon 30 July 2018 Generating Varied Object Shapes and Subcategories With Generative Deep Learning Models

Generating Varied Object Shapes and Subcategories With ... · Generating Varied Object Shapes and Subcategories With Generative Deep Learning Models . Introduction •What’s the

  • Upload
    others

  • View
    4

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Generating Varied Object Shapes and Subcategories With ... · Generating Varied Object Shapes and Subcategories With Generative Deep Learning Models . Introduction •What’s the

Bing Yu (COMP8755)Supervisor: Jo Plested, Tom Gedeon30 July 2018

Generating Varied Object Shapes and Subcategories With Generative Deep Learning Models

Page 2: Generating Varied Object Shapes and Subcategories With ... · Generating Varied Object Shapes and Subcategories With Generative Deep Learning Models . Introduction •What’s the

Introduction

• What’s the problem?– Shape conditioned image generation

2

Page 3: Generating Varied Object Shapes and Subcategories With ... · Generating Varied Object Shapes and Subcategories With Generative Deep Learning Models . Introduction •What’s the

Introduction

• Motivation– In Image Fixing task, available information is limited– Many previous works on conditional image generation – But using shape as condition is rare

3

Page 4: Generating Varied Object Shapes and Subcategories With ... · Generating Varied Object Shapes and Subcategories With Generative Deep Learning Models . Introduction •What’s the

Background

• GAN• Variational Autoencoder• PixelCNN

4

Page 5: Generating Varied Object Shapes and Subcategories With ... · Generating Varied Object Shapes and Subcategories With Generative Deep Learning Models . Introduction •What’s the

Approach

• PixelCNN

5

Page 6: Generating Varied Object Shapes and Subcategories With ... · Generating Varied Object Shapes and Subcategories With Generative Deep Learning Models . Introduction •What’s the

Approach

• Shape conditioned PixelCNN

6

Shape Vector

Conditional PixelCNN

Page 7: Generating Varied Object Shapes and Subcategories With ... · Generating Varied Object Shapes and Subcategories With Generative Deep Learning Models . Introduction •What’s the

Preliminary Result

• Dataset selection and preprocessing• On polar bear dataset

7

Page 8: Generating Varied Object Shapes and Subcategories With ... · Generating Varied Object Shapes and Subcategories With Generative Deep Learning Models . Introduction •What’s the

Preliminary Result

• Reason and fixing approach– Lack of stochastic in training

• Changing the loss function from MSE to discretized logistic mixture likelihood

– The training image size is too large• Using progressive mode

8

Page 9: Generating Varied Object Shapes and Subcategories With ... · Generating Varied Object Shapes and Subcategories With Generative Deep Learning Models . Introduction •What’s the

RoadmapWeeks Task1 Proving the availability of small dataset on model2~5 Build the progressive growing model6 Running different dataset on modelMidterm Break Finding the baseline and reasonable evaluation method7 Evaluating the model8~10 Modifying and improving11 Writing report, final presentation12 Finishing the report

9

Page 10: Generating Varied Object Shapes and Subcategories With ... · Generating Varied Object Shapes and Subcategories With Generative Deep Learning Models . Introduction •What’s the

Reference:Mr. P Lick Cute Coffee Mug - Best Coffee Mugs http://www.mugs.coffee/coffee-mugs/mr-p-lick-cute-coffee-mug/Viv Champagne Glass https://www.crateandbarrel.com/viv-champagne-glass/s240753Oord, A. V. D., Kalchbrenner, N., & Kavukcuoglu, K. (2016). Pixel recurrent neural networks. arXivpreprint arXiv:1601.06759.

Thank You