Supplementary material: Deep Deformable 3D Caricatures with Learned Shape Control

Yucheol Jung1, Wonjong Jang1, Soongjn Kim1, Jiaolong Yang2, Xin Tong2, Seungyong Lee1

1POSTECH, 2Microsoft Research Asia

This file is a gallery of results from our framework. In section A, We perform reconstructions from 2D caricature landmarks to obtain 3D caricatures. Also, we visually compare results from Alive Caricature [Wu et al. 2018]. In section B, we demonstrate semantic editing on 3D caricatures. In section C, we apply point-handle-based editing on the results of A.

Table of contents [top]

A. 3D caricature reconstruction from 2D landmarks [top]

In this section, we reconstruct 3D caricatures from 2D caricatures by fitting our deformable model to 2D landmarks. For comparison, we also show reconstruction results from Alive Caricature [Wu et al. 2018]. We use publicly available 2D caricature landmark data published by the authors of Alive Caricature (https://github.com/QianyiWu/Caricature-Data). For each 2D caricature, 68 landmarks are annotated. We do not have the rights to redistribute the caricature images corresponding to the 2D landmarks, so we provide file names in the repository for each caricature image instead.

The "2D landmarks" column shows input 2D landmarks (red circle) and corresponding 3D landmarks (blue x) after our reconstruction.

Input image2D landmarksOurs (optimized pose)[Wu et al. 2018] (front)Ours (front)[Wu et al. 2018] (side)Ours (side)
Caricature_w_landmark/1.jpg
Caricature_w_landmark/2.jpg
Caricature_w_landmark/3.jpg
Caricature_w_landmark/4.jpg
Caricature_w_landmark/5.jpg
Caricature_w_landmark/6.jpg
Caricature_w_landmark/7.jpg
Caricature_w_landmark/8.jpg
Caricature_w_landmark/9.jpg
Caricature_w_landmark/10.jpg
Caricature_w_landmark/11.jpg
Caricature_w_landmark/12.jpg
Caricature_w_landmark/13.jpg
Caricature_w_landmark/14.jpg
Caricature_w_landmark/15.jpg
Caricature_w_landmark/16.jpg
Caricature_w_landmark/17.jpg
Caricature_w_landmark/18.jpg
Caricature_w_landmark/19.jpg
Caricature_w_landmark/20.jpg
Caricature_w_landmark/21.jpg
Caricature_w_landmark/22.jpg
Caricature_w_landmark/23.jpg
Caricature_w_landmark/24.jpg
Caricature_w_landmark/25.jpg
Caricature_w_landmark/26.jpg
Caricature_w_landmark/27.jpg
Caricature_w_landmark/28.jpg
Caricature_w_landmark/29.jpg
Caricature_w_landmark/30.jpg
Caricature_w_landmark/31.jpg
Caricature_w_landmark/32.jpg
Caricature_w_landmark/33.jpg
Caricature_w_landmark/34.jpg
Caricature_w_landmark/35.jpg
Caricature_w_landmark/36.jpg
Caricature_w_landmark/37.jpg
Caricature_w_landmark/38.jpg
Caricature_w_landmark/39.jpg
Caricature_w_landmark/40.jpg
Caricature_w_landmark/41.jpg
Caricature_w_landmark/42.jpg
Caricature_w_landmark/43.jpg
Caricature_w_landmark/44.jpg
Caricature_w_landmark/45.jpg
Caricature_w_landmark/46.jpg
Caricature_w_landmark/47.jpg
Caricature_w_landmark/48.jpg
Caricature_w_landmark/49.jpg
Caricature_w_landmark/50.jpg

B. Sementic editing [top]

In this section, we demonstrate semantic editing on three caricatures using "Smile", "BigMouth", "BigNose", "HighCheekbones", "LargeForehead", "RoundFace", "SquareJaw" labels in the WebCariA dataset [Ji et al. 2020].

Example 1

(--)(-)Original(+)(++)
Smile
BigMouth
BigNose
HighCheekbones
LargeForehead
RoundFace
SquareJaw

Example 2

(--)(-)Original(+)(++)
Smile
BigMouth
BigNose
HighCheekbones
LargeForehead
RoundFace
SquareJaw

Example 3

(--)(-)Original(+)(++)
Smile
BigMouth
BigNose
HighCheekbones
LargeForehead
RoundFace
SquareJaw

C. Point-handle-based editing [top]

We present various point-handle-based-editing results on the 3D caricature reconstructions from 2D caricatures. The blue dots represent point handles on the surface. Nose: Pick a point on the nose tip. Move the point to the front. Nose 2: A variation. Chin: Pick a point on the bottom of the chin. Move the chin to the bottom. Chin 2: A variation. Cheek: Pick a point at each side of the cheek. Stretch the two points sideways. Ears: Pick a point at each side of the ears. Stretch the two points.

OriginalNoseNose 2ChinChin 2CheekEars

References