Pop-Out Motion [Kor]

Lee et al. / Pop-Out Motion - 3D-Aware Image Deformation via Learning the Shape Laplacian / CVPR 2022

ģ•ˆė…•ķ•˜ģ„øģš”. ė³ø ķ¬ģŠ¤ķŒ…ģ—ģ„œėŠ” ģ˜¬ķ•“ CVPR에 ė°œķ‘œė  Pop-Out Motionģ“ė¼ėŠ” ė…¼ė¬øģ„ ģ†Œź°œė“œė¦¬ź³ ģž ķ•©ė‹ˆė‹¤. ģžģ—°ģŠ¤ėŸ¬ģš“ 3D-Aware Image Deformationģ„ ģœ„ķ•œ ķ•™ģŠµ źø°ė°˜ģ˜ ķŒŒģ“ķ”„ė¼ģøģ„ ģ œģ•ˆķ•œ ė…¼ė¬øģ“ė©°, 3D Vision, Shape Deformation, 2D-to-3D Reconstruction ė“±ģ˜ ķ‚¤ģ›Œė“œģ— ź“€ģ‹¬ģ“ ģžˆģœ¼ģ‹  ė¶„ė“¤ģ“ė¼ė©“ 논문 본문 ė° ķ”„ė”œģ ķŠø ķŽ˜ģ“ģ§€ė„¼ źµ¬ź²½ķ•“ģ£¼ģ‹œė©“ ź°ģ‚¬ķ•˜ź² ģŠµė‹ˆė‹¤. 핓당 ė…¼ė¬øģ€ ģ œź°€ 1ģ €ģžė”œ ģ°øģ—¬ķ•˜ģ˜€ģœ¼ė©°, KAIST ģ „ģ‚°ķ•™ė¶€ģ˜ ģ„±ėÆ¼ķ˜źµģˆ˜ė‹˜ź³¼ ź¹€ķƒœź· źµģˆ˜ė‹˜ź»˜ģ„œ ģ§€ė„ķ•“ģ£¼ģ…ØģŠµė‹ˆė‹¤. (ģ¢‹ģ€ 연구 ģ§€ė„ė„¼ 핓주신 두 źµģˆ˜ė‹˜ź»˜ ź°ģ‚¬ė“œė¦½ė‹ˆė‹¤.)

1. Problem Definition

ė³ø ė…¼ė¬øģ€ 3D-Aware Image Deformation ģ“ė¼ėŠ” 문제넼 ķ•“ź²°ķ•˜ź³ ģž ķ•©ė‹ˆė‹¤. ģ‚¬ģš©ģžź°€ ģ“ėÆøģ§€ ė‚“ģ˜ ź°ģ²“ ėŖØģ–‘ģ„ ģžģ—°ģŠ¤ėŸ½ź²Œ ė³€ķ˜•ķ•˜ėŠ” ź²ƒģ„ ź°€ėŠ„ķ•˜ź²Œ ķ•˜ė˜, 2D ģ˜ģƒģ˜ 피사첓가 마치 3D 공간에 ģ”“ģž¬ķ•˜ėŠ” 것과 ź°™ģ“ ė³€ķ˜•ķ•  수 ģžˆė„ė” ķ•˜ėŠ” ź²ƒģ“ ėŖ©ķ‘œģž…ė‹ˆė‹¤. ģ“ ė•Œ ģ§ź“€ģ ģø ģ“ėÆøģ§€ ģˆ˜ģ •ģ„ ģœ„ķ•˜ģ—¬ ģ‚¬ģš©ģžź°€ ķ‚¤ķ¬ģøķŠø ė“±ģ˜ Deformation Handle ģ„ ė§¤ź°œģ²“ė”œģ„œ ģ‚¬ģš©ķ•  수 ģžˆė„ė” ķ•©ė‹ˆė‹¤. ģ•„ėž˜ģ˜ 그림 ģ˜ˆģ‹œė„¼ ė³“ģ‹œė©“, ģ‚¬ģš©ģžź°€ ģ“ėÆøģ§€ģ— ķ‚¤ķ¬ģøķŠøė“¤ (ķŒŒėž€ģƒ‰ 원 ķ‘œģ‹œ) ģ„ ģ§€ģ •ķ•˜ź³  ź·ø 중 ķ•˜ė‚˜ė„¼ ģ„ ķƒķ•˜ģ—¬ ģ›€ģ§ģ¼ 경우 (ė¹Øź°„ģƒ‰ ķ™”ģ‚“ķ‘œ ķ‘œģ‹œ), 그림 ė‚“ģ˜ ģ‚¬ėžŒ ź°ģ²“ ėŖØģ–‘ģ“ 그에 ė§žģ¶”ģ–“ ģžģ—°ģŠ¤ėŸ½ź²Œ ė³€ķ˜•ė˜ėŠ” ź²ƒģ„ ė³¼ 수 ģžˆģŠµė‹ˆė‹¤. ģ“ ė•Œ ķŒ”ģ“ 몸통 부분 ģ•žģ— ģœ„ģ¹˜ķ•˜ź²Œ ė˜ź±°ė‚˜, ķ•œ ė°œģ“ 다넸 발 ė’¤ė”œ ź°€ė ¤ģ§€ėŠ” ė“±ģ˜ 3D 공간에 ėŒ€ķ•œ ģ“ķ•“ė„¼ 기반으딜 ķ•œ 영상 ė³€ķ˜•ģ“ ģ¼ģ–“ė‚˜ź²Œ ė©ė‹ˆė‹¤. ģ“ėŸ¬ķ•œ 3D-Aware Image Deformation źø°ėŠ„ģ€ ģøķ„°ė ‰ķ‹°ėøŒ 영상 ķŽøģ§‘ ģ–“ķ”Œė¦¬ģ¼€ģ“ģ…˜ 등에 ģœ ģš©ķ•˜ź²Œ ģ“°ģ¼ 수 ģžˆģŠµė‹ˆė‹¤.

2. Motivation

źø°ģ”“ģ—ė„ 3D 공간에 ėŒ€ķ•œ ģ“ķ•“ė„¼ 기반으딜 영상 ķŽøģ§‘ģ„ ź°€ėŠ„ķ•˜ź²Œ ķ•œ źø°ė²•ė“¤ģ“ ė§Žģ“ ģ—°źµ¬ė˜ģ–“ ģ™”ģ§€ė§Œ, źø€ė”œė²Œķ•œ Scene 정볓 (예. ė·°ķ¬ģøķŠø, ģ¹“ė©”ė¼ ķŒŒė¼ėÆøķ„°, ģ”°ėŖ…) ė‚˜ ź¹Šģ“ 정볓넼 ģˆ˜ģ •ķ•˜ėŠ” ź²ƒģ— ģ œķ•œė˜ģ–“ ģžˆģ—ˆģŠµė‹ˆė‹¤. Human Pose Transfer ģŖ½ģ˜ ģ—°źµ¬ė“¤ģ€ 영상 ģ† ģ‚¬ėžŒģ˜ ģžģ„øė„¼ ė³€ķ˜•ķ•˜ėŠ” ź²ƒģ„ ź°€ėŠ„ķ•˜ź²Œ ķ–ˆģ§€ė§Œ, ģ‚¬ėžŒģ“ ģ•„ė‹Œ 다넸 ģ¢…ė„˜ (예. ė§Œķ™” 캐릭터) ģ˜ 영상 ģ† ź°ģ²“ģ— ėŒ€ķ•“ģ„œėŠ” ė™ģž‘ķ•˜ģ§€ ģ•ŠėŠ”ė‹¤ėŠ” ķ•œź³„ģ ģ“ ģžˆģ—ˆģŠµė‹ˆė‹¤. 3D ėŖØėø 기반 ė³€ķ˜• źø°ė²•ė“¤ģ€ 영상 ģ† ź°ģ²“ ģ¢…ė„˜ģ— źµ­ķ•œė˜ģ§€ ģ•Šź³  ė™ģž‘ķ•œė‹¤ėŠ” ģž„ģ ģ“ ģžˆģ§€ė§Œ, ģž…ė „ ģ˜ģƒģ— ėŒ€ģ‘ė˜ėŠ” ģ •ķ™•ķ•œ 3D ėŖØėøģ„ ķ•„ģš”ė”œ ķ•œė‹¤ėŠ” ė‹Øģ ģ“ ģ”“ģž¬ķ•©ė‹ˆė‹¤. ģ“ėŸ¬ķ•œ ķ•œź³„ģ ė“¤ģ„ ź°œģ„ ķ•˜źø° ģœ„ķ•˜ģ—¬ ģ €ķ¬ ģ—°źµ¬ģ—ģ„œėŠ” ź°ģ²“ ģ¢…ė„˜ģ— źµ­ķ•œė˜ģ§€ ģ•Šź³  ģµœėŒ€ķ•œ ģžģœ ė”­ź²Œ 영상 ė³€ķ˜•ģ“ ź°€ėŠ„ķ•œ ķ”„ė ˆģž„ģ›Œķ¬ė„¼ ź³ ģ•ˆķ•˜ėŠ” ź²ƒģ„ ėŖ©ķ‘œė”œ ķ•˜ģ˜€ģŠµė‹ˆė‹¤.

Idea

ź°ģ²“ ģ¢…ė„˜ģ— źµ­ķ•œė˜ģ§€ ģ•Šź³  ģµœėŒ€ķ•œ ģžģœ ė”­ź²Œ 영상 ė³€ķ˜•ģ“ ź°€ėŠ„ķ•˜ź²Œ ķ•˜źø° ģœ„ķ•˜ģ—¬ ģž…ė „ ģ˜ģƒģœ¼ė”œė¶€ķ„° ė³µģ›ėœ 3D Shape에 ėŒ€ķ•“ Handle-Based Deformation Weight [1] ģ„ 기반으딜 영상 ė³€ķ˜•ģ„ ėŖØėøė§ķ•©ė‹ˆė‹¤. (1) Tetrahedral Mesh ķ˜•ķƒœģ˜ 3D Shape M={V,F}\mathcal{M} = \{\mathcal{V}, \mathcal{F}\} ė° (2) ģ‚¬ģš©ģžź°€ ģ§€ģ •ķ•œ Deformation Handle {Hk}k=1⋯m\{ \mathcal{H}_k \}_{k=1 \cdots m} ģ“ ģ£¼ģ–“ģ”Œģ„ ė•Œ, Handle-Based Deformationģ€ ė‹¤ģŒź³¼ ź°™ģ“ ėŖØėøė§ė©ė‹ˆė‹¤:

vi′=āˆ‘k=1mwk,iTkvi.\mathbf{v}_i' = \sum_{k=1}^{m} w_{k,i} \mathbf{T}_k \mathbf{v}_i.

ģœ„ ģˆ˜ģ‹ģ—ģ„œ vi\mathbf{v}_i 와 vi′\mathbf{v}_i' ėŠ” ģž…ė „ Meshģ˜ ii번째 Vertex에 ėŒ€ķ•œ ė³€ķ˜• ģ „ ė° ė³€ķ˜• 후 ģœ„ģ¹˜, wk,iw_{k,i}ėŠ” Vertex vi\mathbf{v_i}와 Handle Hk\mathcal{H_k}에 ėŒ€ģ‘ė˜ėŠ” Deformation Weight, Tk\mathbf{T_k}ėŠ” ģ‚¬ģš©ģžź°€ Handle Hk\mathcal{H_k} 에 ź°€ķ•˜ėŠ” Affine Transformation ķ–‰ė ¬ģ„ ģ˜ėÆøķ•©ė‹ˆė‹¤.

ģ“ ė•Œ ģ‚¬ģš©ķ•˜ėŠ” Handle-Based Deformation Weight [1] ģ€ ė‹¤ģŒź³¼ ź°™ģ€ ģˆ˜ģ‹ģ„ 통핓 ź³„ģ‚°ė©ė‹ˆė‹¤:

argmin{wk}k=1⋯māˆ‘k=1m12ā€…ā€ŠwkT A wksubjectĀ to:Ā ā€…ā€Šwk,i=1āˆ€is.t.vi∈Hkwk,i=0āˆ€is.t.vi∈Hl,l≠kāˆ‘k=1mwk,i=1,i=1,⋯ ,n,0≤wk,i≤1,k=1,⋯ ,m,i=1,⋯ ,n.\underset{\{ \mathbf{w}_k \}_{k=1 \cdots m}}{\mathop{\mathrm{argmin}}} \sum_{k=1}^{m} \frac{1}{2}\; \mathbf{w}_k^T\, A\, \mathbf{w}_k\\ \text{subject to: }\; w_{k,i} = 1 \quad \forall i \quad \text{s.t.} \quad \mathbf{v}_i \in \mathcal{H}_k \\ \qquad \qquad \qquad w_{k,i} = 0 \quad \forall i \quad \text{s.t.} \quad \mathbf{v}_i\in \mathcal{H}_{l, l \neq k} \\ \qquad \qquad \quad \textstyle \sum_{k=1}^{m} w_{k,i}=1, \enspace i=1,\cdots,n, \\ \qquad \qquad \qquad \qquad \qquad \quad 0 \leq w_{k,i} \leq 1, \enspace k=1,\cdots,m, \enspace i=1,\cdots,n.

ģœ„ ģˆ˜ģ‹ģ—ģ„œ 각 Deformation Handle에 ėŒ€ķ•œ Deformation Weights wk={wk,1,⋯ ,wk,n}T\mathbf{w}_k = \{w_{k,1}, \cdots, w_{k,n}\}^T ėŠ” Deformation Energy AA에 ėŒ€ķ•œ Constrained Optimization ė¬øģ œģ˜ ķ•“ė”œģ„œ ģ •ģ˜ė©ė‹ˆė‹¤.

핓당 Deformation Energy AAėŠ” ģž…ė „ Meshģ˜ Shape Laplacianģ„ ģ“ģš©ķ•˜ģ—¬ ģ •ģ˜ė˜ėŠ”ė°, 2D-to-3D Reconstructionģ„ 통핓 ė³µģ›ėœ Meshė”œė¶€ķ„°ėŠ” ė¶€ģ •ķ™•ķ•œ Shape Laplacianģ“ ź³„ģ‚°ėœė‹¤ėŠ” ė¬øģ œź°€ ģžˆģŠµė‹ˆė‹¤. Shape Laplacianģ€ Mesh Topology (즉, Mesh Vertex ź°„ģ˜ Edgeė”œģ„œ ķ‘œķ˜„ėœ ģ—°ź²° ꓀계) 넼 기반으딜 ķ•˜ģ—¬ ģ •ģ˜ė˜ėŠ”ė°, 2D ģ˜ģƒģœ¼ė”œė¶€ķ„° ģ •ķ™•ķ•œ Mesh Topology 정볓넼 복원할 수 ģžˆėŠ” Topology-Aware Mesh Reconstructionģ€ ģ—¬ėŸ¬ 얓려움들 ė•Œė¬øģ— 아직 풀리지 ģ•Šģ€ 문제딜 ė‚Øģ•„ģžˆģŠµė‹ˆė‹¤. ė”°ė¼ģ„œ ģ €ķ¬ģ˜ 핵심 ģ•„ģ“ė””ģ–“ėŠ” 2Dė”œė¶€ķ„° ė³µģ›ėœ 3D Shape에 ėŒ€ķ•œ Shape Laplacian 정볓넼 ķ•™ģŠµ źø°ė°˜ģ˜ źø°ė²•ģ„ 통핓 ģ •ķ™•ķ•˜ź²Œ ģ˜ˆģø”ķ•œ 후, ģ“ė„¼ Handle-Based Deformation Weight 계산에 ģ“ģš©ķ•˜ėŠ” ź²ƒģž…ė‹ˆė‹¤.

3. Method

ģ•žģ„œ ģ–øźø‰ė“œė øė“Æģ“, ģ €ķ¬ėŠ” 3D-Aware Image Deformationģ„ ėŖØėøė§ķ•˜źø° ģœ„ķ•œ ķ•™ģŠµ źø°ė°˜ģ˜ źø°ė²•ģ„ ģ œģ•ˆķ•©ė‹ˆė‹¤. ģš°ģ„  ģž…ė „ ģ˜ģƒģ— ėŒ€ķ•˜ģ—¬ 3D Reconstruction Method (PIFu [2]) 넼 ģ ģš©ķ•Øģœ¼ė”œģØ 영상 ģ† ź°ģ²“ģ— ėŒ€ģ‘ķ•˜ėŠ” 3D Point Cloud넼 ģ˜ˆģø”ķ•©ė‹ˆė‹¤. (ģ €ķ¬ėŠ” Mesh Edge 정볓가 ģ‚¬ģš©ė˜ėŠ” Shape Laplacian ź³„ģ‚°ģ„ ķ•™ģŠµ źø°ė°˜ģ˜ źø°ė²•ģœ¼ė”œ ėŒ€ģ²“ķ•  ź²ƒģ“źø° ė•Œė¬øģ—, Meshź°€ ģ•„ė‹Œ Point Cloud ķ˜•ķƒœģ˜ Shapeģ„ ģ‚¬ģš©ķ•©ė‹ˆė‹¤.) ė‹¤ģŒģ€ ė³µģ›ėœ 3D Point Cloud에 ėŒ€ķ•œ Shape Laplacianģ„ ģ„øģ‹¬ķ•˜ź²Œ ģ„¤ź³„ėœ ė‰“ėŸ“ė„·ģ„ ģ“ģš©ķ•˜ģ—¬ ģ˜ˆģø”ķ•©ė‹ˆė‹¤. ģ“ė ‡ź²Œ 예츔된 Shape Laplacianģ„ ģ“ģš©ķ•˜ģ—¬ ģ‚¬ģš©ģžź°€ ģž„ģ˜ė”œ ģ§€ģ •ķ•œ Deformation Handle에 ėŒ€ķ•œ Handle-Based Deformation Weight [1]ģ„ ź³„ģ‚°ķ•˜ź³ , ģ“ė„¼ 통핓 ėŖØėøė§ 된 3D Deformationģ„ ė‹¤ģ‹œ 2D Image Plane에 ķˆ¬ģ‚¬ķ•Øģœ¼ė”œģØ 3D-Aware Image Deformationģ„ ź°€ėŠ„ķ•˜ź²Œ ķ•©ė‹ˆė‹¤.

ģ§€źøˆė¶€ķ„°ėŠ” ģ €ķ¬ģ˜ 핵심 ģ•„ģ“ė””ģ–“ģø Point Cloudė”œė¶€ķ„° Shape Laplacianģ„ ģ˜ˆģø”ķ•˜ėŠ” ė„¤ķŠøģ›Œķ¬ģ— ėŒ€ķ•˜ģ—¬ ģžģ„øķ•˜ź²Œ ģ†Œź°œė“œė¦¬ź² ģŠµė‹ˆė‹¤. Shape Laplacianģ˜ 구성 ģš”ģ†Œģø Cotangent Laplacian Matrix L∈RnƗnL \in \mathbb{R}^{n \times n} 와 Inverse Mass Matrix Māˆ’1∈RnƗnM^{-1} \in \mathbb{R}^{n \times n} 넼 ė”°ė”œ ģ˜ˆģø”ķ•˜ė„ė” ė„¤ķŠøģ›Œķ¬ė„¼ źµ¬ģ„±ķ•œ 후, 각 정볓에 ėŒ€ķ•œ ģ§ģ ‘ģ ģø Superivsionģ„ ķ†µķ•˜ģ—¬ ė„¤ķŠøģ›Œķ¬ė„¼ ķ•™ģŠµģ‹œķ‚µė‹ˆė‹¤. ģ•„ėž˜ģ˜ ź·øė¦¼ģ„ ė³“ģ‹œė©“ ģ•Œ 수 ģžˆė“Æģ“, ģ œģ•ˆ ķ”„ė ˆģž„ģ›Œķ¬ėŠ” 크게 세 ź°€ģ§€ģ˜ ėŖØė“ˆ - (1) Feature Extraction Module, (2) Cotangent Laplacian Prediction Module, (3) Inverse Mass Prediction Module - 딜 źµ¬ģ„±ė˜ģ–“ģžˆģŠµė‹ˆė‹¤.

Feature Extraction Moduleģ€ ģž…ė „ 2D ģ“ėÆøģ§€ė”œė¶€ķ„° ė³µģ›ėœ 3D Point Cloud P={pi}i=1⋯n\mathcal{P} = \{ \mathbf{p}_i \}_{i = 1 \cdots n} 넼 ģž…ė „ģœ¼ė”œ 받아 Point Cloud Feature F={fi}i=1⋯n\mathcal{F} = \{ \mathbf{f}_i \}_{i = 1 \cdots n} 넼 ģƒģ„±ķ•©ė‹ˆė‹¤. ģ“ ė•Œ fi∈Rd\mathbf{f}_i \in \mathbb{R} ^ d ģ€ pi\mathbf{p}_i 에 ėŒ€ģ‘ė˜ėŠ” Per-Point Feature넼 ģ˜ėÆøķ•©ė‹ˆė‹¤. ėŖØė“ˆģ˜ źµ¬ģ”°ė”œėŠ” Point Transformer [3] 넼 ķ™œģš©ķ•˜ģ˜€ģŠµė‹ˆė‹¤.

Cotangent Laplacian Prediction Moduleģ€ 3D Point Cloud P={pi}i=1⋯n\mathcal{P} = \{ \mathbf{p}_i \}_{i = 1 \cdots n} 와 Point Cloud Feature F={fi}i=1⋯n\mathcal{F} = \{ \mathbf{f}_i \}_{i = 1 \cdots n} 넼 ģž…ė „ģœ¼ė”œ 받아 P\mathcal{P}에 ėŒ€ķ•œ Cotangent Laplacian Matrix L∈RnƗnL \in \mathbb{R}^{n \times n} 넼 ģ˜ˆģø”ķ•©ė‹ˆė‹¤. Cotangent Laplacianģ˜ ģ •ģ˜ģ— ė”°ė¼ LLģ€ Symmetricķ•˜ź³  매우 Sparseķ•œ ķŠ¹ģ„±ģ„ 가지고 ģžˆėŠ”ė°, pi\mathbf{p}_i 와 pj\mathbf{p}_j ģ‚¬ģ“ģ˜ Edge ģ—°ź²° ꓀계가 ģžˆģ–“ģ•¼ LijL_{ij}ģ“ 0ģ“ ģ•„ė‹Œ ź°’ģœ¼ė”œ ģ •ģ˜ė˜źø° ė•Œė¬øģž…ė‹ˆė‹¤. ģ €ķ¬ėŠ” Point Cloud ė‚“ģ˜ 각 Point Pair (pi\mathbf{p}_i, pj\mathbf{p}_j) 넼 ģž…ė „ģœ¼ė”œ 받아 ģ“ģ— ėŒ€ģ‘ė˜ėŠ” Laplacian Matrixģ˜ Element (LijL_{ij}) 넼 ė³‘ė ¬ģ ģœ¼ė”œ ģ˜ˆģø”ķ•˜ėŠ” 구씰넼 ģ·Øķ•˜ėŠ”ė°, Euclidean Distanceź°€ 먼 Point Pair ė¼ė¦¬ėŠ” ģ—°ź²° ꓀계가 ģžˆģ„ ķ™•ė„ ģ“ 적기 ė•Œė¬øģ— ģ“ė“¤ģ„ 1차적으딜 ź±øėŸ¬ģ£¼ėŠ” ģ—­ķ• ģ„ ķ•©ė‹ˆė‹¤. ė…¼ė¬øģ—ģ„œ KNN-Based Point Pair Sampling (KPS) 으딜 ģ§€ģ¹­ķ•˜ėŠ” ė¶€ė¶„ģøė°, 각 ķ¬ģøķŠøė“¤ģ— ėŒ€ķ•˜ģ—¬ kk ź°œģ˜ ź°€ź¹Œģš“ 점들에 ėŒ€ķ•“ģ„œė§Œ Point Pair넼 źµ¬ģ„±ķ•˜ėŠ” źø°ė²•ģž…ė‹ˆė‹¤. ģ“ėŸ¬ķ•œ Sampling źø°ė²•ģ„ ģ“°ģ§€ ģ•Šģ„ 경우 Imbalanced Regression Problemģ“ ģ¼ģ–“ė‚˜ ė„¤ķŠøģ›Œķ¬ ķ•™ģŠµģ“ ģž˜ ė˜ģ§€ ģ•ŠėŠ” ķ˜„ģƒģ“ ģžˆģ—ˆģŠµė‹ˆė‹¤.

ė‹¤ģŒģ€ KNN-Based Point Pair Sampling (KPS) ģ„ 통핓 ģ„ ķƒėœ 각 Point Pair Candidate (pi\mathbf{p}_i, pj\mathbf{p}_j) 에 ėŒ€ķ•˜ģ—¬ Symmetric Feature Aggregation ģ„ ģˆ˜ķ–‰ķ•“ģ¤ė‹ˆė‹¤:

gm=(γ1(pi, pj),γ2(fi,fj)).\mathbf{g}_{m} = ( \gamma_1(\mathbf{p}_i,\, \mathbf{p}_j), \gamma_2(\mathbf{f}_i, \mathbf{f}_j) ).

ģœ„ ģˆ˜ģ‹ģ—ģ„œ γ1(ā‹…)\gamma_1(\cdot) ė° γ2(ā‹…)\gamma_2(\cdot)ė”œėŠ” Symmetric Functionģ„ ģ‚¬ģš©ķ•˜ėŠ”ė°, ģ“ėŠ” ė‚˜ģ¤‘ģ— 예츔될 Cotangent Laplacian Matrixģ˜ Symmetry넼 ė³“ģž„ķ•˜źø° ģœ„ķ•Øģž…ė‹ˆė‹¤. 핓당 ķ•Øģˆ˜ėŠ” 각각 Absolute Difference와 Element-Wise Multiplication으딜 źµ¬ķ˜„ė˜ģ—ˆģŠµė‹ˆė‹¤. ģ“ė ‡ź²Œ ģƒģ„±ėœ Point Pair Feature gm\mathbf{g}_{m} 에 ėŒ€ģ‘ė˜ėŠ” Cotangent Laplacian Element LijL_{ij}ėŠ” ė‹¤ģŒź³¼ ź°™ģ“ ģ˜ˆģø”ė©ė‹ˆė‹¤:

Lij=α(gm)āŠ™Ļ•(gm).L_{ij} = \alpha(\mathbf{g}_{m}) \odot \phi(\mathbf{g}_{m}).

Ļ•(ā‹…)\phi(\cdot) ģ€ Real-Valued Scalar넼 ģ¶œė „ķ•˜ėŠ” ķ•Øģˆ˜ģ“ė©° α(ā‹…)\alpha(\cdot) ėŠ” LijL_{ij}ģ“ Non-Zero ź°’ģ¼ģ§€ģ— ėŒ€ķ•œ ķ™•ė„ ģ„ ėŖØėøė§ķ•˜ėŠ” Weight Wij∈[0,1]W_{ij} \in [0, 1] 출렄 ķ•Øģˆ˜ģž…ė‹ˆė‹¤. 두 ķ•Øģˆ˜ėŠ” MLP딜 źµ¬ķ˜„ė˜ģ—ˆģœ¼ė©°, ģµœģ¢… LijL_{ij} ź°’ģ€ 두 출렄 ź°’ģ˜ ź³±ģœ¼ė”œģ„œ ķ‘œķ˜„ė©ė‹ˆė‹¤.

Inverse Mass Prediction Moduleģ€ 3D Point Cloud P={pi}i=1⋯n\mathcal{P} = \{ \mathbf{p}_i \}_{i = 1 \cdots n} 와 Point Cloud Feature F={fi}i=1⋯n\mathcal{F} = \{ \mathbf{f}_i \}_{i = 1 \cdots n} 넼 ģž…ė „ģœ¼ė”œ 받아 P\mathcal{P}에 ėŒ€ķ•œ Inverse Mass Matrix Māˆ’1∈RnƗnM^{-1} \in \mathbb{R}^{n \times n} 넼 ģ˜ˆģø”ķ•©ė‹ˆė‹¤. Inverse Massģ˜ ģ •ģ˜ģ— ė”°ė¼ Māˆ’1M^{-1} ģ€ Diagonal ķ•˜ė©°, ii번째 Digonal ElementėŠ” pi\mathbf{p}_iģ˜ Volumeź³¼ ź“€ź³„ėœ 정볓넼 ė‹“ź³  ģžˆģŠµė‹ˆė‹¤. ė”°ė¼ģ„œ P\mathcal{P} ė‚“ģ˜ 각 ķ¬ģøķŠø pi\mathbf{p}_i 와 ėŒ€ģ‘ė˜ėŠ” Per-Point Feature fi\mathbf{f}_i 넼 Concatenate ģ‹œģ¼œģ¤€ 후 MLP에 ķ†µź³¼ģ‹œķ‚¤ėŠ” ė°©ģ‹ģ„ 통핓 Inverse Mass Matrix ė‚“ģ˜ Miiāˆ’1M^{-1}_{ii} Element넼 ģ˜ˆģø”ķ•©ė‹ˆė‹¤.

ė³ø Shape Laplacian 예츔 ė„¤ķŠøģ›Œķ¬ėŠ” LL, WW, Māˆ’1M^{-1} 예츔 값에 ėŒ€ķ•œ L1-Loss źø°ė°˜ģ˜ Ground Truth Supervisionģ„ 통핓 ķ•™ģŠµė©ė‹ˆė‹¤. ģžģ„øķ•œ Loss 계산 ģ •ė³“ėŠ” 논문 ė³øė¬øģ„ ģ°øģ”°ķ•“ģ£¼ģ‹œė©“ ź°ģ‚¬ķ•˜ź² ģŠµė‹ˆė‹¤.

4. Experiment & Result

ģ œģ•ˆķ•œ 3D-Aware Image Deformation źø°ė²•ģ˜ ķšØź³¼ģ„±ģ„ ź²€ģ¦ķ•˜źø° ģœ„ķ•˜ģ—¬ 크게 두 ģ¢…ė„˜ģ˜ ģ‹¤ķ—˜ģ„ ģ§„ķ–‰ķ•˜ģ˜€ģŠµė‹ˆė‹¤. 첫 ė²ˆģ§øė”œėŠ” ģ €ķ¬ź°€ ėŖØėøė§ķ•œ Deformationģ˜ 퀄리티넼 ģ •ėŸ‰ģ ģœ¼ė”œ ķ‰ź°€ķ•˜źø° ģœ„ķ•“ 3D Point Cloud Deformation ģ‹¤ķ—˜ģ„ ģ§„ķ–‰ķ•˜ģ˜€ģŠµė‹ˆė‹¤. 두 ė²ˆģ§øė”œėŠ” ģ €ķ¬ģ˜ ėŖ©ķ‘œ źø°ėŠ„ģø 3D-Aware Image Deformation 결과넼 ķ™•ģøķ•˜źø° ģœ„ķ•œ 정성적 ķ‰ź°€ė„¼ ģ§„ķ–‰ķ•˜ģ˜€ģŠµė‹ˆė‹¤. ė”ģš± ė‹¤ģ–‘ķ•œ ģ‹¤ķ—˜ ź²°ź³¼ (예. Partial Point Cloud Deformation, Ablation Study) ėŠ” 논문 ė³øė¬øģ—ģ„œ ķ™•ģøķ•“ģ£¼ģ‹œė©“ ź°ģ‚¬ķ•˜ź² ģŠµė‹ˆė‹¤.

Experimental setup

  • Dataset

    • DFAUST [4]: ģ •ėŸ‰ģ  ķ‰ź°€ģ— ģ‚¬ģš©ėœ 3D Human Point Cloud ė°ģ“ķ„°ģ…‹ģž…ė‹ˆė‹¤.

    • RenderPeople [5], Mixamo [6]: 정성적 ķ‰ź°€ģ— ģ‚¬ģš©ėœ 3D Human [5] ė° 3D Character [6] Datasetģž…ė‹ˆė‹¤. ģ €ķ¬ģ˜ ėŖ©ģ ģ€ Image Deformationģ˜ 결과넼 ķ™•ģøķ•˜ėŠ” ź²ƒģ“ėÆ€ė”œ, 핓당 3D Modelė“¤ģ„ ė Œė”ė§ķ•˜ģ—¬ ģƒģ„±ķ•œ ģ˜ģƒė“¤ģ„ ģ‹¤ķ—˜ģ— ģ‚¬ģš©ķ•˜ģ˜€ģŠµė‹ˆė‹¤.

  • Baselines

    • ģ €ķ¬ģ˜ 핵심 ģ•„ģ“ė””ģ–“ėŠ” Mesh Reconstruction ź²°ź³¼ė”œė¶€ķ„° ė¶€ģ •ķ™•ķ•œ Shape Laplacianģ“ ź³„ģ‚°ė˜ėÆ€ė”œ 핓당 정볓넼 ķ•™ģŠµ źø°ė°˜ģ˜ źø°ė²•ģ„ 통핓 볓다 ģ •ķ™•ķ•˜ź²Œ ģ˜ˆģø”ķ•˜ģžėŠ” ź²ƒģ“ģ—ˆģŠµė‹ˆė‹¤. ė”°ė¼ģ„œ, Mesh Reconstruction źø°ė²•ģ„ ģ‚¬ģš©ķ•˜ģ—¬ Shape Laplacianģ„ ģ–»ģ€ 후 Deformation Weightģ„ ź³„ģ‚°ķ•˜ėŠ” ģƒķ™©ģ„ ė² ģ“ģŠ¤ė¼ģøģœ¼ė”œ ģ„¤ģ •ķ•˜ģ˜€ģŠµė‹ˆė‹¤. ģ €ķ¬ ģ‹¤ķ—˜ģ—ģ„œ 고려된 Mesh Reconstruction źø°ė²•ė“¤ģ€ ė‹¤ģŒź³¼ ź°™ģŠµė‹ˆė‹¤:

      • Screened Poisson Surface Reconstruction (PSR) [7],

      • Algebraic Point Set Surfaces (APSS) [8],

      • Ball-Pivoting Algorithm (BPA) [9],

      • DeepSDF [10],

      • Deep Geometric Prior (DGP) [11],

      • Meshing Point Clouds with IntrinsicExtrinsic Ratio (MIER) [12].

    • ė˜ķ•œ, źø°ģ”“ģ˜ Point Cloud Laplacian źø°ė²•ģ„ ģ“ģš©ķ•˜ģ—¬ ģž…ė „ Point Cloudė”œė¶€ķ„° Shape Laplacianģ˜ 근사 ź°’ģ„ ė°”ė”œ ź³„ģ‚°ķ•˜ėŠ” źø°ė²•ė“¤ė„ ź³ ė ¤ķ•˜ģ˜€ģŠµė‹ˆė‹¤:

      • PCD Laplace (PCDLap) [13],

      • Nonmanifold Laplacians (NMLap) [14].

  • Training Setup

    • 각 ė°ģ“ķ„°ė³„ė”œ ģ‹¤ķ—˜ģ— ģ‚¬ģš©ķ•œ ģ„øķŒ…ģ“ ė‹¤ė„“ėÆ€ė”œ, ģžģ„øķ•œ ģ‚¬ķ•­ģ€ 논문 본문 ė° Supplementary넼 ģ°øź³ ķ•“ģ£¼ģ‹œė©“ ź°ģ‚¬ķ•˜ź² ģŠµė‹ˆė‹¤.

  • Evaluation Metric

    • ģ €ķ¬ģ˜ ģ •ėŸ‰ģ  ķ‰ź°€ģ—ėŠ” ė‹¤ģŒź³¼ ź°™ģ€ ė©”ķŠøė¦­ģ“ ģ‚¬ģš©ė˜ģ—ˆģŠµė‹ˆė‹¤:

      • 예츔 ė° 정답 Deformation Weights ź°„ģ˜ L1 Distance (Weight L1),

      • 예츔 ė° 정답 Deformed Shape ź°„ģ˜ Chamfer Distance (Shape CD),

      • 예츔 ė° 정답 Deformed Shape ź°„ģ˜ Hausdorff Distance (Shape HD).

Result

3D Point Cloud Deformation

ģ•„ėž˜ģ˜ ķ‘œėŠ” DFAUST [4] ė°ģ“ķ„°ģ…‹ģ— ėŒ€ķ•œ ģ •ėŸ‰ģ  비교 ķ‰ź°€ 결과넼 ė‚˜ķƒ€ė‚ø ź²ƒģž…ė‹ˆė‹¤. ģ €ķ¬ź°€ ģ œģ•ˆķ•œ źø°ė²•ģ“ 다넸 Mesh Reconstruction ė² ģ“ģŠ¤ė¼ģø źø°ė²•ė“¤ģ„ ģ‚¬ģš©ķ–ˆģ„ ė•Œ 볓다 ė” ė‚˜ģ€ Shape Deformation 결과넼 ė³“ģ“ėŠ” ź²ƒģ„ ģ•Œ 수 ģžˆģŠµė‹ˆė‹¤.

ģœ„ģ˜ 결과에 ėŒ€ķ•œ 정성적 ź²°ź³¼ (ģ•„ėž˜ 그림) ė˜ķ•œ ģ €ķ¬ źø°ė²•ģ“ ė”ģš± ģžģ—°ģŠ¤ėŸ¬ģš“ Shape Deformationģ„ ėŖØėøė§ķ•  수 ģžˆģŒģ„ ė³“ģ—¬ģ¤ė‹ˆė‹¤.

3D-Aware Image Deformation

ė³ø ė™ģ˜ģƒģ€ ģ €ķ¬ģ˜ 3D-Aware Image Deformation źø°ė²•ģ„ ģ“ģš©ķ•“ģ„œ ģƒģ„±ķ•œ ėŖØģ…˜ ė™ģ˜ģƒģž…ė‹ˆė‹¤. Mesh Reconstruction ė² ģ“ģŠ¤ė¼ģø 기법들볓다 ė”ģš± ģžģ—°ģŠ¤ėŸ¬ģš“ Image Deformationģ„ ģƒģ„±ķ•  수 ģžˆģŒģ„ ė³“ģ—¬ģ¤ė‹ˆė‹¤.

Interactive Demoė„ ģ²“ķ—˜ķ•“ė³“ģ‹œźø°ė„¼ ė°”ėžė‹ˆė‹¤. ģ‚¬ģš©ģžź°€ ģ§ź“€ģ ģø Deformation Handle (Keypoint)넼 ģ“ģš©ķ•˜ģ—¬ ģ˜ģƒģ„ ė³€ķ˜•ķ•  수 ģžˆģŠµė‹ˆė‹¤.

5. Conclusion

ė³ø ģ—°źµ¬ģ—ģ„œėŠ” Shape Laplacianģ„ ķ•™ģŠµķ•Øģœ¼ė”œģØ 볓다 ģžģ—°ģŠ¤ėŸ¬ģš“ 3D-Aware Deformationģ„ ź°€ėŠ„ķ•˜ź²Œķ•˜ėŠ” ķ”„ė ˆģž„ģ›Œķ¬ė„¼ ģ œģ•ˆķ•˜ģ˜€ģŠµė‹ˆė‹¤. ģ €ķ¬ź°€ ģ•Œźø°ė”œėŠ” ģ“ź°€ ė‰“ėŸ“ė„· 기반 źø°ė²•ģ“ Shape Lapacian ģ˜ˆģø”ģ— ķšØź³¼ģ ģ¼ 수 ģžˆģŒģ„ ģ²˜ģŒģœ¼ė”œ ė³“ģø ģ—°źµ¬ė¼ź³  ģ•Œź³  ģžˆģŠµė‹ˆė‹¤. ė³ø ķ”„ė ˆģž„ģ›Œķ¬ė„¼ ė°œģ „ģ‹œķ‚¤źø° ģœ„ķ•œ ė”ģš± ė‹¤ģ–‘ķ•œ ģ•„ģ“ė””ģ–“ź°€ ė§Žģ€ė°, źø°ķšŒź°€ ėœė‹¤ė©“ 핓당 ė°©ķ–„ģœ¼ė”œ ė”ģš± 연구핓볓고 ģ‹¶ģŠµė‹ˆė‹¤.

Take-Home Message (ģ˜¤ėŠ˜ģ˜ źµķ›ˆ)

ģ œź°€ ź°œģøģ ģœ¼ė”œ ģ“ ķ”„ė”œģ ķŠøė„¼ 통핓 배욓 źµķ›ˆģ€ "ėź¹Œģ§€ ķ¬źø°ķ•˜ģ§€ ģ•Šź³  ģ§‘ė…ģ„ 가지며 연구 문제넼 ķ’€ģž"ėŠ” ź²ƒģž…ė‹ˆė‹¤. ė³ø ķ”„ė ˆģž„ģ›Œķ¬ 개발 ė‹Øź³„ģ—ģ„œ ģžģž˜ķ•œ Challengeė“¤ģ“ ė§Žģ•˜ģ—ˆź³ , ź·ø 과정 중 ģ›ėž˜ ģ§„ķ–‰ķ•˜ė ¤ė˜ 연구 ė°©ķ–„ģœ¼ė”œė¶€ķ„° 크게 ė°”ė€Œģ–“ 마묓리된 ė¶€ė¶„ė„ ģžˆģŠµė‹ˆė‹¤. ź·øėž˜ė„ ė™ģž‘ķ•˜ėŠ” ģ†”ė£Øģ…˜ģ„ 찾아낓고 뜻 깊게 ķ”„ė”œģ ķŠøė„¼ ė§ˆė¬“ė¦¬ķ•  수 ģžˆģ–“ģ„œ ź°œģøģ ģœ¼ė”œėŠ” 매우 기억에 ė‚ØėŠ” 연구 ź²½ķ—˜ģ“ ė˜ģ—ˆģŠµė‹ˆė‹¤. ź·ø 과정 중 큰 ė„ģ›€ź³¼ ģ”°ģ–øģ„ 주신 두 ģ§€ė„ źµģˆ˜ė‹˜ź»˜ ź¹Šģ€ 감사넼 ė“œė¦½ė‹ˆė‹¤.

Author / Reviewer information

Author

ģ“ģ§€ķ˜„ (Jihyun Lee)

  • KAIST CS

  • I am a first-year Ph.D. student in Computer Vision and Learning Lab at KAIST advised by Prof. Tae-Kyun Kim. I am also currently co-advised by Prof. Minhyuk Sung. My research interests lie in machine learning for 3D computer vision and graphics - especially on humans.

Reviewer

  1. Korean name (English name): Affiliation / Contact information

  2. Korean name (English name): Affiliation / Contact information

  3. ...

Reference & Additional materials

  1. Citation of related work

    1. Alec Jacobson, Ilya Baran, Jovan Popovic, and Olga Sorkine. Bounded biharmonic weights for real-time deformation. In SIGGRAPH, 2011.

    2. Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, 2019.

    3. Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021.

    4. Federica Bogo, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Dynamic FAUST: Registering human bodies in motion. In CVPR, 2017.

    5. Renderpeople GmbH. RenderPeople. https://renderpeople.com/.

    6. Adobe Systems Inc. Mixamo. https://www.mixamo.com.

    7. Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM TOG, 2013.

    8. Gael Guennebaud and Markus Gross. Algebraic point set surfaces. In SIGGRAPH, 2007.

    9. Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Claudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. IEEE TVCG, 1999.

    10. Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In CVPR, 2019.

    11. Francis Williams, Teseo Schneider, Claudio Silva, Denis Zorin, Joan Bruna, and Daniele Panozzo. Deep geometric prior for surface reconstruction. In CVPR, 2019.

    12. Minghua Liu, Xiaoshuai Zhang, and Hao Su. Meshing point clouds with predicted intrinsic-extrinsic ratio guidance. In ECCV, 2020.

    13. Mikhail Belkin, Jian Sun, and Yusu Wang. Constructing laplace operator from point clouds in rd. In Proc. Annu. ACM-SIAM Symp. Discrete Algorithms, pages 1031–1040. SIAM, 2009.

    14. Nicholas Sharp and Keenan Crane. A laplacian for nonmanifold triangle meshes. In SGP, 2020.

Last updated

Was this helpful?