[논문]Predicting 3D body shape and body composition from conventional 2D photography

2020 [ body composition from 2D photo ]

0. Inputs

• weight & height (for 3D shape initialization)
• DXA body composition
• GT 3D optical scans
• 2D photographys

---

Estimate DXA body composition from 3D optical scan

---

Detecting joint location & Segmenting subject from 2D photo

---

Fitting mesh template to 3D optical scan W/ PCA

• by initializing template mesh with W&H + ridig transformation(T)

[ Template fitting ]

[ PCA ]

• eigen decomposition(고유값 분해)로 PCA basis 구함
• PCA basis : d col vertors (n = 180 003)
  ; 3D를 1D로 이어붙여서(flatten?) 고윳값 분해(?) - 60 001 3D(x.y.z) -> 180 003 1D

• M(뮤) = 기본 신체 형태 (template mesh)
• A = PCA 주성분
• w = PCA coefficient (mean과의 차이 : 이 계수 조정해 body shape 바꿈)
  +) 80 PCA vectors가 99% 설명함 -> d = 80 이라는 의미

---

Fitting 3D body mesh(#2) to silhouette from 2D

• by minimizing energy(E) function

---

Mapping 3D PCA coefficients to body composition

---

### Predicting expected body compositon(#1) from fitted 3D shape(#3)