[DL Hacks]Smart, Sparse Contours to Represent and Edit Images

【CVPR2018論文紹介 】 Sparse, Smart Contours to Represent and Edit Images 2018/07/02 DLHacks Otsubo

Topic1 : Image Encoding and Decoding (Code = Sparse Contour + α) 1

Topic2 : Image Editing 2

Info • Sparse, Smart Contours to Represent and Edit Image [CVPR2018] - https://contour2im.github.io/ - https://arxiv.org/abs/1712.08232 3

• https://contour2im.github.io/
• https://arxiv.org/abs/1712.08232

選んだ理由 • Contourからの画像復元とか無理では？ あまりにもill-posed → 少量の情報（勾配）を追加するだけで実現 • Encoding Decodingにおいて 人間が理解，編集しやすく 機械が復元しやすいCodeを模索 • イケメンになりたい 4

Encoder Overview 1. Extract Edge [1] 2. Get Contour (remove less than 10pixel-connected) 3. Compute addiConal channel
[1] Structured forests for fast edge detecCon [Dollar, ICCV2013]
5

Encoder Additional Channel 1. Color - (Rleft, Gleft, Bleft, Rright, Gright, Bright ) 2. Gradient 編集に最適(reconstructionでは弱い) - (Rdx, Gdx, Bdx, Rdy, Gdy, Bdy ) 3. Learned Feature - Semantic Segmentationのように学習 6

Decoder Overview 7

Decoder Train (1/2) L1 Loss
8

Decoder Train (2/2) L1 Loss + adv loss (a)
9

Reconstruction Result 10

Reconstruction Result FaceNet Style Transfer  "! = Texture  11

Reconstruction Result ReconstrucConGradient 12

Editing 13

Online Demo https://contours2im.appspot.com/ 14

• https://contours2im.appspot.com/