Review SketchNet: Sketch Classification with Web Images [CVPR `16] - PowerPoint PPT Presentation

Review ● SketchNet: Sketch Classification with Web Images [CVPR `16] (Speaker. Doheon Lee) ● Problem in previous sketch-based image retrieval ● People have different sketch style ● Large difference btw sketch and image ● Manual Annotation is expensive ● Solution ● Weakly supervised Learning ● Triplet pair (anchor sketch, positive & negative images) ● Sketch Net: S-Net (sketch), R-Net (image), and C-Net ● C-Net: merge feature maps btw image and sketch 1

Age Progression/Regression by Conditional Adversarial Autoencoder [CVPR `17] 20189008 Ben Jung ( 정병의 )

Table of Contents ● Introduction ● Problems of Previous Works ● Main Idea & Solution: CAAE ● Experiment & Result ● Overview 3

Introduction ● Age Progression & Regression Regression Progression/Aging Given face 10 20 35 years old 40 50 4

Problems of Previous Works ● Group-wised learning ● Query with label ● Step-by-step transition 5 years old 10 30 60 … … ! " ! $ ! # query with label 10 5

Main Idea ● Group-wised learning è Joint learning ● Query with label è Query without label ● Step-by-step transition è One-step & bidirectional transition query 6

Main Idea: Manifold Traversing ● Assumptions ● The faces lies on a manifold ( ! ) ● Clustered by ages and personality ● Traversing on the manifold corresponds to age/personality transformation 7

Solution: CAAE ● Conditional Adversarial Autoencoder 8

Solution: CAAE ● Conditional Adversarial Autoencoder ! #$% ! " Uniform noise Real faces 9

Solution: CAAE ● Effect of Discriminator on z ( ! " ) 10

Solution: CAAE ● Effect of Discriminator on image ( ! "#$ ) without ! "#$ with ! "#$ without ! "#$ with ! "#$ without ! "#$ with ! "#$ 11

Experiment & Result 12

Experiment & Result ● Comparison with prior work 13

Experiment & Result ● Comparison with ground truth 14

Overview ● CAAE (Conditional Adversarial Autoencoder) ● Manifold Traversing ● Joint learning ● Query without label ● One-step & bidirectional transition ● Discriminator on z ● Discriminator on image 15

THANK YOU 16