for 3 d c ont r ol l a bl e i m a ge s y nt h e s i s towa r ds u ns upe … · 2020-05-09 · towa...
TRANSCRIPT
![Page 1: for 3 D C ont r ol l a bl e I m a ge S y nt h e s i s Towa r ds U ns upe … · 2020-05-09 · Towa r ds U ns upe r v i s e d Le a r ni n g o f G e n e r a t i v e M o d e l s for](https://reader034.vdocuments.us/reader034/viewer/2022043020/5f3c03a385ae8c3f7654c74c/html5/thumbnails/1.jpg)
Towards Unsupervised Learning of Generative Models for 3D Controllable Image Synthesis
Yiyi Liao*, Katja Schwarz*, Lars Mescheder, Andreas Geiger
![Page 2: for 3 D C ont r ol l a bl e I m a ge S y nt h e s i s Towa r ds U ns upe … · 2020-05-09 · Towa r ds U ns upe r v i s e d Le a r ni n g o f G e n e r a t i v e M o d e l s for](https://reader034.vdocuments.us/reader034/viewer/2022043020/5f3c03a385ae8c3f7654c74c/html5/thumbnails/2.jpg)
Motivation3D Controllable Image Synthesis
3D controllability is essential in many applications, e.g., gaming, simulation, virtualreality and data augmentation3D controllable properties: 3D pose, shape, appearance of multiple objects andcamera viewpoint
![Page 3: for 3 D C ont r ol l a bl e I m a ge S y nt h e s i s Towa r ds U ns upe … · 2020-05-09 · Towa r ds U ns upe r v i s e d Le a r ni n g o f G e n e r a t i v e M o d e l s for](https://reader034.vdocuments.us/reader034/viewer/2022043020/5f3c03a385ae8c3f7654c74c/html5/thumbnails/3.jpg)
MotivationClassical Rendering Pipeline
Render3D Design
3D ControllableExpensive and inefficient to design 3D models
![Page 4: for 3 D C ont r ol l a bl e I m a ge S y nt h e s i s Towa r ds U ns upe … · 2020-05-09 · Towa r ds U ns upe r v i s e d Le a r ni n g o f G e n e r a t i v e M o d e l s for](https://reader034.vdocuments.us/reader034/viewer/2022043020/5f3c03a385ae8c3f7654c74c/html5/thumbnails/4.jpg)
Motivation2D Generative Models
2D Generator
Sample
Efficient, learned from only 2D imagesGeometry and appearance not disentangled → no 3D control
![Page 5: for 3 D C ont r ol l a bl e I m a ge S y nt h e s i s Towa r ds U ns upe … · 2020-05-09 · Towa r ds U ns upe r v i s e d Le a r ni n g o f G e n e r a t i v e M o d e l s for](https://reader034.vdocuments.us/reader034/viewer/2022043020/5f3c03a385ae8c3f7654c74c/html5/thumbnails/5.jpg)
MotivationOurs
3D Generator
Render2D
GeneratorSample
3D ControllableEfficient, learned from only 2D imagesUnsupervised, disentangled 3D representation learning
![Page 6: for 3 D C ont r ol l a bl e I m a ge S y nt h e s i s Towa r ds U ns upe … · 2020-05-09 · Towa r ds U ns upe r v i s e d Le a r ni n g o f G e n e r a t i v e M o d e l s for](https://reader034.vdocuments.us/reader034/viewer/2022043020/5f3c03a385ae8c3f7654c74c/html5/thumbnails/6.jpg)
MotivationOurs
3D Generator
Render2D
GeneratorSample
Idea: Learning the image synthesis pipeline jointly in 3D and 2D space
![Page 7: for 3 D C ont r ol l a bl e I m a ge S y nt h e s i s Towa r ds U ns upe … · 2020-05-09 · Towa r ds U ns upe r v i s e d Le a r ni n g o f G e n e r a t i v e M o d e l s for](https://reader034.vdocuments.us/reader034/viewer/2022043020/5f3c03a385ae8c3f7654c74c/html5/thumbnails/7.jpg)
Method Overview
3D Generator Differentiable Rendering 2D Generator
Render
Render
Render
Render
Alpha
Composition
Real
Background
Image
BackgroundImage
Real
![Page 8: for 3 D C ont r ol l a bl e I m a ge S y nt h e s i s Towa r ds U ns upe … · 2020-05-09 · Towa r ds U ns upe r v i s e d Le a r ni n g o f G e n e r a t i v e M o d e l s for](https://reader034.vdocuments.us/reader034/viewer/2022043020/5f3c03a385ae8c3f7654c74c/html5/thumbnails/8.jpg)
ExperimentsDatasets
Car w/o BG Car with BG Indoor Fruit
![Page 9: for 3 D C ont r ol l a bl e I m a ge S y nt h e s i s Towa r ds U ns upe … · 2020-05-09 · Towa r ds U ns upe r v i s e d Le a r ni n g o f G e n e r a t i v e M o d e l s for](https://reader034.vdocuments.us/reader034/viewer/2022043020/5f3c03a385ae8c3f7654c74c/html5/thumbnails/9.jpg)
ExperimentsCar Dataset
Layout2Im and our 2D baseline areonly controllable for 2D translationLayout2Im fails to disentangle objectidentity and poseOur method is controllable for 3Dtranslation and rotation withcoherent object identity
![Page 10: for 3 D C ont r ol l a bl e I m a ge S y nt h e s i s Towa r ds U ns upe … · 2020-05-09 · Towa r ds U ns upe r v i s e d Le a r ni n g o f G e n e r a t i v e M o d e l s for](https://reader034.vdocuments.us/reader034/viewer/2022043020/5f3c03a385ae8c3f7654c74c/html5/thumbnails/10.jpg)
ExperimentsIndoor Dataset
Layout2Im and our 2D baseline areonly controllable for 2D translationLayout2Im fails to disentangle objectidentity and poseOur method is controllable for 3Dtranslation and rotation withcoherent object identity
![Page 11: for 3 D C ont r ol l a bl e I m a ge S y nt h e s i s Towa r ds U ns upe … · 2020-05-09 · Towa r ds U ns upe r v i s e d Le a r ni n g o f G e n e r a t i v e M o d e l s for](https://reader034.vdocuments.us/reader034/viewer/2022043020/5f3c03a385ae8c3f7654c74c/html5/thumbnails/11.jpg)
ExperimentsFruit Dataset
We collect 800 images with 5 fruitsand 5 backgroundsOur method is able to synthesizeplausible images from real data
![Page 12: for 3 D C ont r ol l a bl e I m a ge S y nt h e s i s Towa r ds U ns upe … · 2020-05-09 · Towa r ds U ns upe r v i s e d Le a r ni n g o f G e n e r a t i v e M o d e l s for](https://reader034.vdocuments.us/reader034/viewer/2022043020/5f3c03a385ae8c3f7654c74c/html5/thumbnails/12.jpg)
Failure Cases
A single primitive generates multiple objects occasionally
Identity might flip wrt. large viewpoint change
Stronger inductive biases are required to tackle these problems
![Page 13: for 3 D C ont r ol l a bl e I m a ge S y nt h e s i s Towa r ds U ns upe … · 2020-05-09 · Towa r ds U ns upe r v i s e d Le a r ni n g o f G e n e r a t i v e M o d e l s for](https://reader034.vdocuments.us/reader034/viewer/2022043020/5f3c03a385ae8c3f7654c74c/html5/thumbnails/13.jpg)
Thank you!