Pop-Out Motion [Kor]

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

์•ˆ๋…•ํ•˜์„ธ์š”. ๋ณธ ํฌ์ŠคํŒ…์—์„œ๋Š” ์˜ฌํ•ด CVPR์— ๋ฐœํ‘œ๋  Pop-Out Motion์ด๋ผ๋Š” ๋…ผ๋ฌธ์„ ์†Œ๊ฐœ๋“œ๋ฆฌ๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค. ์ž์—ฐ์Šค๋Ÿฌ์šด 3D-Aware Image Deformation์„ ์œ„ํ•œ ํ•™์Šต ๊ธฐ๋ฐ˜์˜ ํŒŒ์ดํ”„๋ผ์ธ์„ ์ œ์•ˆํ•œ ๋…ผ๋ฌธ์ด๋ฉฐ, 3D Vision, Shape Deformation, 2D-to-3D Reconstruction ๋“ฑ์˜ ํ‚ค์›Œ๋“œ์— ๊ด€์‹ฌ์ด ์žˆ์œผ์‹  ๋ถ„๋“ค์ด๋ผ๋ฉด ๋…ผ๋ฌธ ๋ณธ๋ฌธ ๋ฐ ํ”„๋กœ์ ํŠธ ํŽ˜์ด์ง€๋ฅผ ๊ตฌ๊ฒฝํ•ด์ฃผ์‹œ๋ฉด ๊ฐ์‚ฌํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. ํ•ด๋‹น ๋…ผ๋ฌธ์€ ์ œ๊ฐ€ 1์ €์ž๋กœ ์ฐธ์—ฌํ•˜์˜€์œผ๋ฉฐ, KAIST ์ „์‚ฐํ•™๋ถ€์˜ ์„ฑ๋ฏผํ˜๊ต์ˆ˜๋‹˜๊ณผ ๊น€ํƒœ๊ท ๊ต์ˆ˜๋‹˜๊ป˜์„œ ์ง€๋„ํ•ด์ฃผ์…จ์Šต๋‹ˆ๋‹ค. (์ข‹์€ ์—ฐ๊ตฌ ์ง€๋„๋ฅผ ํ•ด์ฃผ์‹  ๋‘ ๊ต์ˆ˜๋‹˜๊ป˜ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค.)

1. Problem Definition

๋ณธ ๋…ผ๋ฌธ์€ 3D-Aware Image Deformation ์ด๋ผ๋Š” ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ์šฉ์ž๊ฐ€ ์ด๋ฏธ์ง€ ๋‚ด์˜ ๊ฐ์ฒด ๋ชจ์–‘์„ ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ๋ณ€ํ˜•ํ•˜๋Š” ๊ฒƒ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๋˜, 2D ์˜์ƒ์˜ ํ”ผ์‚ฌ์ฒด๊ฐ€ ๋งˆ์น˜ 3D ๊ณต๊ฐ„์— ์กด์žฌํ•˜๋Š” ๊ฒƒ๊ณผ ๊ฐ™์ด ๋ณ€ํ˜•ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ์ž…๋‹ˆ๋‹ค. ์ด ๋•Œ ์ง๊ด€์ ์ธ ์ด๋ฏธ์ง€ ์ˆ˜์ •์„ ์œ„ํ•˜์—ฌ ์‚ฌ์šฉ์ž๊ฐ€ ํ‚คํฌ์ธํŠธ ๋“ฑ์˜ Deformation Handle ์„ ๋งค๊ฐœ์ฒด๋กœ์„œ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ์•„๋ž˜์˜ ๊ทธ๋ฆผ ์˜ˆ์‹œ๋ฅผ ๋ณด์‹œ๋ฉด, ์‚ฌ์šฉ์ž๊ฐ€ ์ด๋ฏธ์ง€์— ํ‚คํฌ์ธํŠธ๋“ค (ํŒŒ๋ž€์ƒ‰ ์› ํ‘œ์‹œ) ์„ ์ง€์ •ํ•˜๊ณ  ๊ทธ ์ค‘ ํ•˜๋‚˜๋ฅผ ์„ ํƒํ•˜์—ฌ ์›€์ง์ผ ๊ฒฝ์šฐ (๋นจ๊ฐ„์ƒ‰ ํ™”์‚ดํ‘œ ํ‘œ์‹œ), ๊ทธ๋ฆผ ๋‚ด์˜ ์‚ฌ๋žŒ ๊ฐ์ฒด ๋ชจ์–‘์ด ๊ทธ์— ๋งž์ถ”์–ด ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ๋ณ€ํ˜•๋˜๋Š” ๊ฒƒ์„ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋•Œ ํŒ”์ด ๋ชธํ†ต ๋ถ€๋ถ„ ์•ž์— ์œ„์น˜ํ•˜๊ฒŒ ๋˜๊ฑฐ๋‚˜, ํ•œ ๋ฐœ์ด ๋‹ค๋ฅธ ๋ฐœ ๋’ค๋กœ ๊ฐ€๋ ค์ง€๋Š” ๋“ฑ์˜ 3D ๊ณต๊ฐ„์— ๋Œ€ํ•œ ์ดํ•ด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์˜์ƒ ๋ณ€ํ˜•์ด ์ผ์–ด๋‚˜๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ 3D-Aware Image Deformation ๊ธฐ๋Šฅ์€ ์ธํ„ฐ๋ ‰ํ‹ฐ๋ธŒ ์˜์ƒ ํŽธ์ง‘ ์–ดํ”Œ๋ฆฌ์ผ€์ด์…˜ ๋“ฑ์— ์œ ์šฉํ•˜๊ฒŒ ์“ฐ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

2. Motivation

๊ธฐ์กด์—๋„ 3D ๊ณต๊ฐ„์— ๋Œ€ํ•œ ์ดํ•ด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์˜์ƒ ํŽธ์ง‘์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•œ ๊ธฐ๋ฒ•๋“ค์ด ๋งŽ์ด ์—ฐ๊ตฌ๋˜์–ด ์™”์ง€๋งŒ, ๊ธ€๋กœ๋ฒŒํ•œ Scene ์ •๋ณด (์˜ˆ. ๋ทฐํฌ์ธํŠธ, ์นด๋ฉ”๋ผ ํŒŒ๋ผ๋ฏธํ„ฐ, ์กฐ๋ช…) ๋‚˜ ๊นŠ์ด ์ •๋ณด๋ฅผ ์ˆ˜์ •ํ•˜๋Š” ๊ฒƒ์— ์ œํ•œ๋˜์–ด ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. Human Pose Transfer ์ชฝ์˜ ์—ฐ๊ตฌ๋“ค์€ ์˜์ƒ ์† ์‚ฌ๋žŒ์˜ ์ž์„ธ๋ฅผ ๋ณ€ํ˜•ํ•˜๋Š” ๊ฒƒ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ–ˆ์ง€๋งŒ, ์‚ฌ๋žŒ์ด ์•„๋‹Œ ๋‹ค๋ฅธ ์ข…๋ฅ˜ (์˜ˆ. ๋งŒํ™” ์บ๋ฆญํ„ฐ) ์˜ ์˜์ƒ ์† ๊ฐ์ฒด์— ๋Œ€ํ•ด์„œ๋Š” ๋™์ž‘ํ•˜์ง€ ์•Š๋Š”๋‹ค๋Š” ํ•œ๊ณ„์ ์ด ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. 3D ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ๋ณ€ํ˜• ๊ธฐ๋ฒ•๋“ค์€ ์˜์ƒ ์† ๊ฐ์ฒด ์ข…๋ฅ˜์— ๊ตญํ•œ๋˜์ง€ ์•Š๊ณ  ๋™์ž‘ํ•œ๋‹ค๋Š” ์žฅ์ ์ด ์žˆ์ง€๋งŒ, ์ž…๋ ฅ ์˜์ƒ์— ๋Œ€์‘๋˜๋Š” ์ •ํ™•ํ•œ 3D ๋ชจ๋ธ์„ ํ•„์š”๋กœ ํ•œ๋‹ค๋Š” ๋‹จ์ ์ด ์กด์žฌํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ•œ๊ณ„์ ๋“ค์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•˜์—ฌ ์ €ํฌ ์—ฐ๊ตฌ์—์„œ๋Š” ๊ฐ์ฒด ์ข…๋ฅ˜์— ๊ตญํ•œ๋˜์ง€ ์•Š๊ณ  ์ตœ๋Œ€ํ•œ ์ž์œ ๋กญ๊ฒŒ ์˜์ƒ ๋ณ€ํ˜•์ด ๊ฐ€๋Šฅํ•œ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๊ณ ์•ˆํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜์˜€์Šต๋‹ˆ๋‹ค.

Idea

๊ฐ์ฒด ์ข…๋ฅ˜์— ๊ตญํ•œ๋˜์ง€ ์•Š๊ณ  ์ตœ๋Œ€ํ•œ ์ž์œ ๋กญ๊ฒŒ ์˜์ƒ ๋ณ€ํ˜•์ด ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•˜์—ฌ ์ž…๋ ฅ ์˜์ƒ์œผ๋กœ๋ถ€ํ„ฐ ๋ณต์›๋œ 3D Shape์— ๋Œ€ํ•ด Handle-Based Deformation Weight [1] ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์˜์ƒ ๋ณ€ํ˜•์„ ๋ชจ๋ธ๋งํ•ฉ๋‹ˆ๋‹ค. (1) Tetrahedral Mesh ํ˜•ํƒœ์˜ 3D Shape M={V,F}\mathcal{M} = \{\mathcal{V}, \mathcal{F}\} ๋ฐ (2) ์‚ฌ์šฉ์ž๊ฐ€ ์ง€์ •ํ•œ Deformation Handle {Hk}k=1โ‹ฏm\{ \mathcal{H}_k \}_{k=1 \cdots m} ์ด ์ฃผ์–ด์กŒ์„ ๋•Œ, Handle-Based Deformation์€ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๋ชจ๋ธ๋ง๋ฉ๋‹ˆ๋‹ค:

viโ€ฒ=โˆ‘k=1mwk,iTkvi.\mathbf{v}_i' = \sum_{k=1}^{m} w_{k,i} \mathbf{T}_k \mathbf{v}_i.

์œ„ ์ˆ˜์‹์—์„œ vi\mathbf{v}_i ์™€ viโ€ฒ\mathbf{v}_i' ๋Š” ์ž…๋ ฅ Mesh์˜ ii๋ฒˆ์งธ Vertex์— ๋Œ€ํ•œ ๋ณ€ํ˜• ์ „ ๋ฐ ๋ณ€ํ˜• ํ›„ ์œ„์น˜, wk,iw_{k,i}๋Š” Vertex vi\mathbf{v_i}์™€ Handle Hk\mathcal{H_k}์— ๋Œ€์‘๋˜๋Š” Deformation Weight, Tk\mathbf{T_k}๋Š” ์‚ฌ์šฉ์ž๊ฐ€ Handle Hk\mathcal{H_k} ์— ๊ฐ€ํ•˜๋Š” Affine Transformation ํ–‰๋ ฌ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.

์ด ๋•Œ ์‚ฌ์šฉํ•˜๋Š” Handle-Based Deformation Weight [1] ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ˆ˜์‹์„ ํ†ตํ•ด ๊ณ„์‚ฐ๋ฉ๋‹ˆ๋‹ค:

argmin{wk}k=1โ‹ฏmโˆ‘k=1m12โ€…โ€ŠwkTโ€‰Aโ€‰wksubjectย to:ย โ€…โ€Šwk,i=1โˆ€is.t.viโˆˆHkwk,i=0โˆ€is.t.viโˆˆHl,lโ‰ kโˆ‘k=1mwk,i=1,i=1,โ‹ฏโ€‰,n,0โ‰คwk,iโ‰ค1,k=1,โ‹ฏโ€‰,m,i=1,โ‹ฏโ€‰,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.

์œ„ ์ˆ˜์‹์—์„œ ๊ฐ Deformation Handle์— ๋Œ€ํ•œ Deformation Weights wk={wk,1,โ‹ฏโ€‰,wk,n}T\mathbf{w}_k = \{w_{k,1}, \cdots, w_{k,n}\}^T ๋Š” Deformation Energy AA์— ๋Œ€ํ•œ Constrained Optimization ๋ฌธ์ œ์˜ ํ•ด๋กœ์„œ ์ •์˜๋ฉ๋‹ˆ๋‹ค.

ํ•ด๋‹น Deformation Energy AA๋Š” ์ž…๋ ฅ Mesh์˜ Shape Laplacian์„ ์ด์šฉํ•˜์—ฌ ์ •์˜๋˜๋Š”๋ฐ, 2D-to-3D Reconstruction์„ ํ†ตํ•ด ๋ณต์›๋œ Mesh๋กœ๋ถ€ํ„ฐ๋Š” ๋ถ€์ •ํ™•ํ•œ Shape Laplacian์ด ๊ณ„์‚ฐ๋œ๋‹ค๋Š” ๋ฌธ์ œ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. Shape Laplacian์€ Mesh Topology (์ฆ‰, Mesh Vertex ๊ฐ„์˜ Edge๋กœ์„œ ํ‘œํ˜„๋œ ์—ฐ๊ฒฐ ๊ด€๊ณ„) ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜์—ฌ ์ •์˜๋˜๋Š”๋ฐ, 2D ์˜์ƒ์œผ๋กœ๋ถ€ํ„ฐ ์ •ํ™•ํ•œ Mesh Topology ์ •๋ณด๋ฅผ ๋ณต์›ํ•  ์ˆ˜ ์žˆ๋Š” Topology-Aware Mesh Reconstruction์€ ์—ฌ๋Ÿฌ ์–ด๋ ค์›€๋“ค ๋•Œ๋ฌธ์— ์•„์ง ํ’€๋ฆฌ์ง€ ์•Š์€ ๋ฌธ์ œ๋กœ ๋‚จ์•„์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ €ํฌ์˜ ํ•ต์‹ฌ ์•„์ด๋””์–ด๋Š” 2D๋กœ๋ถ€ํ„ฐ ๋ณต์›๋œ 3D Shape์— ๋Œ€ํ•œ Shape Laplacian ์ •๋ณด๋ฅผ ํ•™์Šต ๊ธฐ๋ฐ˜์˜ ๊ธฐ๋ฒ•์„ ํ†ตํ•ด ์ •ํ™•ํ•˜๊ฒŒ ์˜ˆ์ธกํ•œ ํ›„, ์ด๋ฅผ Handle-Based Deformation Weight ๊ณ„์‚ฐ์— ์ด์šฉํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.

3. Method

์•ž์„œ ์–ธ๊ธ‰๋“œ๋ ธ๋“ฏ์ด, ์ €ํฌ๋Š” 3D-Aware Image Deformation์„ ๋ชจ๋ธ๋งํ•˜๊ธฐ ์œ„ํ•œ ํ•™์Šต ๊ธฐ๋ฐ˜์˜ ๊ธฐ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์šฐ์„  ์ž…๋ ฅ ์˜์ƒ์— ๋Œ€ํ•˜์—ฌ 3D Reconstruction Method (PIFu [2]) ๋ฅผ ์ ์šฉํ•จ์œผ๋กœ์จ ์˜์ƒ ์† ๊ฐ์ฒด์— ๋Œ€์‘ํ•˜๋Š” 3D Point Cloud๋ฅผ ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค. (์ €ํฌ๋Š” Mesh Edge ์ •๋ณด๊ฐ€ ์‚ฌ์šฉ๋˜๋Š” Shape Laplacian ๊ณ„์‚ฐ์„ ํ•™์Šต ๊ธฐ๋ฐ˜์˜ ๊ธฐ๋ฒ•์œผ๋กœ ๋Œ€์ฒดํ•  ๊ฒƒ์ด๊ธฐ ๋•Œ๋ฌธ์—, Mesh๊ฐ€ ์•„๋‹Œ Point Cloud ํ˜•ํƒœ์˜ Shape์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.) ๋‹ค์Œ์€ ๋ณต์›๋œ 3D Point Cloud์— ๋Œ€ํ•œ Shape Laplacian์„ ์„ธ์‹ฌํ•˜๊ฒŒ ์„ค๊ณ„๋œ ๋‰ด๋Ÿด๋„ท์„ ์ด์šฉํ•˜์—ฌ ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ์˜ˆ์ธก๋œ Shape Laplacian์„ ์ด์šฉํ•˜์—ฌ ์‚ฌ์šฉ์ž๊ฐ€ ์ž„์˜๋กœ ์ง€์ •ํ•œ Deformation Handle์— ๋Œ€ํ•œ Handle-Based Deformation Weight [1]์„ ๊ณ„์‚ฐํ•˜๊ณ , ์ด๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ๋ง ๋œ 3D Deformation์„ ๋‹ค์‹œ 2D Image Plane์— ํˆฌ์‚ฌํ•จ์œผ๋กœ์จ 3D-Aware Image Deformation์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค.

์ง€๊ธˆ๋ถ€ํ„ฐ๋Š” ์ €ํฌ์˜ ํ•ต์‹ฌ ์•„์ด๋””์–ด์ธ Point Cloud๋กœ๋ถ€ํ„ฐ Shape Laplacian์„ ์˜ˆ์ธกํ•˜๋Š” ๋„คํŠธ์›Œํฌ์— ๋Œ€ํ•˜์—ฌ ์ž์„ธํ•˜๊ฒŒ ์†Œ๊ฐœ๋“œ๋ฆฌ๊ฒ ์Šต๋‹ˆ๋‹ค. Shape Laplacian์˜ ๊ตฌ์„ฑ ์š”์†Œ์ธ Cotangent Laplacian Matrix LโˆˆRnร—nL \in \mathbb{R}^{n \times n} ์™€ Inverse Mass Matrix Mโˆ’1โˆˆRnร—nM^{-1} \in \mathbb{R}^{n \times n} ๋ฅผ ๋”ฐ๋กœ ์˜ˆ์ธกํ•˜๋„๋ก ๋„คํŠธ์›Œํฌ๋ฅผ ๊ตฌ์„ฑํ•œ ํ›„, ๊ฐ ์ •๋ณด์— ๋Œ€ํ•œ ์ง์ ‘์ ์ธ Superivsion์„ ํ†ตํ•˜์—ฌ ๋„คํŠธ์›Œํฌ๋ฅผ ํ•™์Šต์‹œํ‚ต๋‹ˆ๋‹ค. ์•„๋ž˜์˜ ๊ทธ๋ฆผ์„ ๋ณด์‹œ๋ฉด ์•Œ ์ˆ˜ ์žˆ๋“ฏ์ด, ์ œ์•ˆ ํ”„๋ ˆ์ž„์›Œํฌ๋Š” ํฌ๊ฒŒ ์„ธ ๊ฐ€์ง€์˜ ๋ชจ๋“ˆ - (1) Feature Extraction Module, (2) Cotangent Laplacian Prediction Module, (3) Inverse Mass Prediction Module - ๋กœ ๊ตฌ์„ฑ๋˜์–ด์žˆ์Šต๋‹ˆ๋‹ค.

Feature Extraction Module์€ ์ž…๋ ฅ 2D ์ด๋ฏธ์ง€๋กœ๋ถ€ํ„ฐ ๋ณต์›๋œ 3D Point Cloud P={pi}i=1โ‹ฏn\mathcal{P} = \{ \mathbf{p}_i \}_{i = 1 \cdots n} ๋ฅผ ์ž…๋ ฅ์œผ๋กœ ๋ฐ›์•„ Point Cloud Feature F={fi}i=1โ‹ฏn\mathcal{F} = \{ \mathbf{f}_i \}_{i = 1 \cdots n} ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด ๋•Œ fiโˆˆRd\mathbf{f}_i \in \mathbb{R} ^ d ์€ pi\mathbf{p}_i ์— ๋Œ€์‘๋˜๋Š” Per-Point Feature๋ฅผ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“ˆ์˜ ๊ตฌ์กฐ๋กœ๋Š” Point Transformer [3] ๋ฅผ ํ™œ์šฉํ•˜์˜€์Šต๋‹ˆ๋‹ค.

Cotangent Laplacian Prediction Module์€ 3D Point Cloud P={pi}i=1โ‹ฏn\mathcal{P} = \{ \mathbf{p}_i \}_{i = 1 \cdots n} ์™€ Point Cloud Feature F={fi}i=1โ‹ฏn\mathcal{F} = \{ \mathbf{f}_i \}_{i = 1 \cdots n} ๋ฅผ ์ž…๋ ฅ์œผ๋กœ ๋ฐ›์•„ P\mathcal{P}์— ๋Œ€ํ•œ Cotangent Laplacian Matrix LโˆˆRnร—nL \in \mathbb{R}^{n \times n} ๋ฅผ ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค. Cotangent Laplacian์˜ ์ •์˜์— ๋”ฐ๋ผ LL์€ Symmetricํ•˜๊ณ  ๋งค์šฐ Sparseํ•œ ํŠน์„ฑ์„ ๊ฐ€์ง€๊ณ  ์žˆ๋Š”๋ฐ, pi\mathbf{p}_i ์™€ pj\mathbf{p}_j ์‚ฌ์ด์˜ Edge ์—ฐ๊ฒฐ ๊ด€๊ณ„๊ฐ€ ์žˆ์–ด์•ผ LijL_{ij}์ด 0์ด ์•„๋‹Œ ๊ฐ’์œผ๋กœ ์ •์˜๋˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์ €ํฌ๋Š” Point Cloud ๋‚ด์˜ ๊ฐ Point Pair (pi\mathbf{p}_i, pj\mathbf{p}_j) ๋ฅผ ์ž…๋ ฅ์œผ๋กœ ๋ฐ›์•„ ์ด์— ๋Œ€์‘๋˜๋Š” Laplacian Matrix์˜ Element (LijL_{ij}) ๋ฅผ ๋ณ‘๋ ฌ์ ์œผ๋กœ ์˜ˆ์ธกํ•˜๋Š” ๊ตฌ์กฐ๋ฅผ ์ทจํ•˜๋Š”๋ฐ, Euclidean Distance๊ฐ€ ๋จผ Point Pair ๋ผ๋ฆฌ๋Š” ์—ฐ๊ฒฐ ๊ด€๊ณ„๊ฐ€ ์žˆ์„ ํ™•๋ฅ ์ด ์ ๊ธฐ ๋•Œ๋ฌธ์— ์ด๋“ค์„ 1์ฐจ์ ์œผ๋กœ ๊ฑธ๋Ÿฌ์ฃผ๋Š” ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค. ๋…ผ๋ฌธ์—์„œ KNN-Based Point Pair Sampling (KPS) ์œผ๋กœ ์ง€์นญํ•˜๋Š” ๋ถ€๋ถ„์ธ๋ฐ, ๊ฐ ํฌ์ธํŠธ๋“ค์— ๋Œ€ํ•˜์—ฌ kk ๊ฐœ์˜ ๊ฐ€๊นŒ์šด ์ ๋“ค์— ๋Œ€ํ•ด์„œ๋งŒ Point Pair๋ฅผ ๊ตฌ์„ฑํ•˜๋Š” ๊ธฐ๋ฒ•์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ Sampling ๊ธฐ๋ฒ•์„ ์“ฐ์ง€ ์•Š์„ ๊ฒฝ์šฐ Imbalanced Regression Problem์ด ์ผ์–ด๋‚˜ ๋„คํŠธ์›Œํฌ ํ•™์Šต์ด ์ž˜ ๋˜์ง€ ์•Š๋Š” ํ˜„์ƒ์ด ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.

๋‹ค์Œ์€ KNN-Based Point Pair Sampling (KPS) ์„ ํ†ตํ•ด ์„ ํƒ๋œ ๊ฐ Point Pair Candidate (pi\mathbf{p}_i, pj\mathbf{p}_j) ์— ๋Œ€ํ•˜์—ฌ Symmetric Feature Aggregation ์„ ์ˆ˜ํ–‰ํ•ด์ค๋‹ˆ๋‹ค:

gm=(ฮณ1(pi,โ€‰pj),ฮณ2(fi,fj)).\mathbf{g}_{m} = ( \gamma_1(\mathbf{p}_i,\, \mathbf{p}_j), \gamma_2(\mathbf{f}_i, \mathbf{f}_j) ).

์œ„ ์ˆ˜์‹์—์„œ ฮณ1(โ‹…)\gamma_1(\cdot) ๋ฐ ฮณ2(โ‹…)\gamma_2(\cdot)๋กœ๋Š” Symmetric Function์„ ์‚ฌ์šฉํ•˜๋Š”๋ฐ, ์ด๋Š” ๋‚˜์ค‘์— ์˜ˆ์ธก๋  Cotangent Laplacian Matrix์˜ Symmetry๋ฅผ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•จ์ž…๋‹ˆ๋‹ค. ํ•ด๋‹น ํ•จ์ˆ˜๋Š” ๊ฐ๊ฐ Absolute Difference์™€ Element-Wise Multiplication์œผ๋กœ ๊ตฌํ˜„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ์ƒ์„ฑ๋œ Point Pair Feature gm\mathbf{g}_{m} ์— ๋Œ€์‘๋˜๋Š” Cotangent Laplacian Element LijL_{ij}๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์˜ˆ์ธก๋ฉ๋‹ˆ๋‹ค:

Lij=ฮฑ(gm)โŠ™ฯ•(gm).L_{ij} = \alpha(\mathbf{g}_{m}) \odot \phi(\mathbf{g}_{m}).

ฯ•(โ‹…)\phi(\cdot) ์€ Real-Valued Scalar๋ฅผ ์ถœ๋ ฅํ•˜๋Š” ํ•จ์ˆ˜์ด๋ฉฐ ฮฑ(โ‹…)\alpha(\cdot) ๋Š” LijL_{ij}์ด Non-Zero ๊ฐ’์ผ์ง€์— ๋Œ€ํ•œ ํ™•๋ฅ ์„ ๋ชจ๋ธ๋งํ•˜๋Š” Weight Wijโˆˆ[0,1]W_{ij} \in [0, 1] ์ถœ๋ ฅ ํ•จ์ˆ˜์ž…๋‹ˆ๋‹ค. ๋‘ ํ•จ์ˆ˜๋Š” MLP๋กœ ๊ตฌํ˜„๋˜์—ˆ์œผ๋ฉฐ, ์ตœ์ข… LijL_{ij} ๊ฐ’์€ ๋‘ ์ถœ๋ ฅ ๊ฐ’์˜ ๊ณฑ์œผ๋กœ์„œ ํ‘œํ˜„๋ฉ๋‹ˆ๋‹ค.

Inverse Mass Prediction Module์€ 3D Point Cloud P={pi}i=1โ‹ฏn\mathcal{P} = \{ \mathbf{p}_i \}_{i = 1 \cdots n} ์™€ Point Cloud Feature F={fi}i=1โ‹ฏn\mathcal{F} = \{ \mathbf{f}_i \}_{i = 1 \cdots n} ๋ฅผ ์ž…๋ ฅ์œผ๋กœ ๋ฐ›์•„ P\mathcal{P}์— ๋Œ€ํ•œ Inverse Mass Matrix Mโˆ’1โˆˆRnร—nM^{-1} \in \mathbb{R}^{n \times n} ๋ฅผ ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค. Inverse Mass์˜ ์ •์˜์— ๋”ฐ๋ผ Mโˆ’1M^{-1} ์€ Diagonal ํ•˜๋ฉฐ, ii๋ฒˆ์งธ Digonal Element๋Š” pi\mathbf{p}_i์˜ Volume๊ณผ ๊ด€๊ณ„๋œ ์ •๋ณด๋ฅผ ๋‹ด๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ P\mathcal{P} ๋‚ด์˜ ๊ฐ ํฌ์ธํŠธ pi\mathbf{p}_i ์™€ ๋Œ€์‘๋˜๋Š” Per-Point Feature fi\mathbf{f}_i ๋ฅผ Concatenate ์‹œ์ผœ์ค€ ํ›„ MLP์— ํ†ต๊ณผ์‹œํ‚ค๋Š” ๋ฐฉ์‹์„ ํ†ตํ•ด Inverse Mass Matrix ๋‚ด์˜ Miiโˆ’1M^{-1}_{ii} Element๋ฅผ ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค.

๋ณธ Shape Laplacian ์˜ˆ์ธก ๋„คํŠธ์›Œํฌ๋Š” LL, WW, Mโˆ’1M^{-1} ์˜ˆ์ธก ๊ฐ’์— ๋Œ€ํ•œ L1-Loss ๊ธฐ๋ฐ˜์˜ Ground Truth Supervision์„ ํ†ตํ•ด ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. ์ž์„ธํ•œ Loss ๊ณ„์‚ฐ ์ •๋ณด๋Š” ๋…ผ๋ฌธ ๋ณธ๋ฌธ์„ ์ฐธ์กฐํ•ด์ฃผ์‹œ๋ฉด ๊ฐ์‚ฌํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.

4. Experiment & Result

์ œ์•ˆํ•œ 3D-Aware Image Deformation ๊ธฐ๋ฒ•์˜ ํšจ๊ณผ์„ฑ์„ ๊ฒ€์ฆํ•˜๊ธฐ ์œ„ํ•˜์—ฌ ํฌ๊ฒŒ ๋‘ ์ข…๋ฅ˜์˜ ์‹คํ—˜์„ ์ง„ํ–‰ํ•˜์˜€์Šต๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ๋กœ๋Š” ์ €ํฌ๊ฐ€ ๋ชจ๋ธ๋งํ•œ Deformation์˜ ํ€„๋ฆฌํ‹ฐ๋ฅผ ์ •๋Ÿ‰์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด 3D Point Cloud Deformation ์‹คํ—˜์„ ์ง„ํ–‰ํ•˜์˜€์Šต๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ๋กœ๋Š” ์ €ํฌ์˜ ๋ชฉํ‘œ ๊ธฐ๋Šฅ์ธ 3D-Aware Image Deformation ๊ฒฐ๊ณผ๋ฅผ ํ™•์ธํ•˜๊ธฐ ์œ„ํ•œ ์ •์„ฑ์  ํ‰๊ฐ€๋ฅผ ์ง„ํ–‰ํ•˜์˜€์Šต๋‹ˆ๋‹ค. ๋”์šฑ ๋‹ค์–‘ํ•œ ์‹คํ—˜ ๊ฒฐ๊ณผ (์˜ˆ. Partial Point Cloud Deformation, Ablation Study) ๋Š” ๋…ผ๋ฌธ ๋ณธ๋ฌธ์—์„œ ํ™•์ธํ•ด์ฃผ์‹œ๋ฉด ๊ฐ์‚ฌํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.

Experimental setup

  • Dataset

    • DFAUST [4]: ์ •๋Ÿ‰์  ํ‰๊ฐ€์— ์‚ฌ์šฉ๋œ 3D Human Point Cloud ๋ฐ์ดํ„ฐ์…‹์ž…๋‹ˆ๋‹ค.

    • RenderPeople [5], Mixamo [6]: ์ •์„ฑ์  ํ‰๊ฐ€์— ์‚ฌ์šฉ๋œ 3D Human [5] ๋ฐ 3D Character [6] Dataset์ž…๋‹ˆ๋‹ค. ์ €ํฌ์˜ ๋ชฉ์ ์€ Image Deformation์˜ ๊ฒฐ๊ณผ๋ฅผ ํ™•์ธํ•˜๋Š” ๊ฒƒ์ด๋ฏ€๋กœ, ํ•ด๋‹น 3D Model๋“ค์„ ๋ Œ๋”๋งํ•˜์—ฌ ์ƒ์„ฑํ•œ ์˜์ƒ๋“ค์„ ์‹คํ—˜์— ์‚ฌ์šฉํ•˜์˜€์Šต๋‹ˆ๋‹ค.

  • Baselines

    • ์ €ํฌ์˜ ํ•ต์‹ฌ ์•„์ด๋””์–ด๋Š” Mesh Reconstruction ๊ฒฐ๊ณผ๋กœ๋ถ€ํ„ฐ ๋ถ€์ •ํ™•ํ•œ Shape Laplacian์ด ๊ณ„์‚ฐ๋˜๋ฏ€๋กœ ํ•ด๋‹น ์ •๋ณด๋ฅผ ํ•™์Šต ๊ธฐ๋ฐ˜์˜ ๊ธฐ๋ฒ•์„ ํ†ตํ•ด ๋ณด๋‹ค ์ •ํ™•ํ•˜๊ฒŒ ์˜ˆ์ธกํ•˜์ž๋Š” ๊ฒƒ์ด์—ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ, Mesh Reconstruction ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ Shape Laplacian์„ ์–ป์€ ํ›„ Deformation Weight์„ ๊ณ„์‚ฐํ•˜๋Š” ์ƒํ™ฉ์„ ๋ฒ ์ด์Šค๋ผ์ธ์œผ๋กœ ์„ค์ •ํ•˜์˜€์Šต๋‹ˆ๋‹ค. ์ €ํฌ ์‹คํ—˜์—์„œ ๊ณ ๋ ค๋œ Mesh Reconstruction ๊ธฐ๋ฒ•๋“ค์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:

      • 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].

    • ๋˜ํ•œ, ๊ธฐ์กด์˜ Point Cloud Laplacian ๊ธฐ๋ฒ•์„ ์ด์šฉํ•˜์—ฌ ์ž…๋ ฅ Point Cloud๋กœ๋ถ€ํ„ฐ Shape Laplacian์˜ ๊ทผ์‚ฌ ๊ฐ’์„ ๋ฐ”๋กœ ๊ณ„์‚ฐํ•˜๋Š” ๊ธฐ๋ฒ•๋“ค๋„ ๊ณ ๋ คํ•˜์˜€์Šต๋‹ˆ๋‹ค:

      • PCD Laplace (PCDLap) [13],

      • Nonmanifold Laplacians (NMLap) [14].

  • Training Setup

    • ๊ฐ ๋ฐ์ดํ„ฐ๋ณ„๋กœ ์‹คํ—˜์— ์‚ฌ์šฉํ•œ ์„ธํŒ…์ด ๋‹ค๋ฅด๋ฏ€๋กœ, ์ž์„ธํ•œ ์‚ฌํ•ญ์€ ๋…ผ๋ฌธ ๋ณธ๋ฌธ ๋ฐ Supplementary๋ฅผ ์ฐธ๊ณ ํ•ด์ฃผ์‹œ๋ฉด ๊ฐ์‚ฌํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.

  • Evaluation Metric

    • ์ €ํฌ์˜ ์ •๋Ÿ‰์  ํ‰๊ฐ€์—๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋ฉ”ํŠธ๋ฆญ์ด ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค:

      • ์˜ˆ์ธก ๋ฐ ์ •๋‹ต Deformation Weights ๊ฐ„์˜ L1 Distance (Weight L1),

      • ์˜ˆ์ธก ๋ฐ ์ •๋‹ต Deformed Shape ๊ฐ„์˜ Chamfer Distance (Shape CD),

      • ์˜ˆ์ธก ๋ฐ ์ •๋‹ต Deformed Shape ๊ฐ„์˜ Hausdorff Distance (Shape HD).

Result

3D Point Cloud Deformation

์•„๋ž˜์˜ ํ‘œ๋Š” DFAUST [4] ๋ฐ์ดํ„ฐ์…‹์— ๋Œ€ํ•œ ์ •๋Ÿ‰์  ๋น„๊ต ํ‰๊ฐ€ ๊ฒฐ๊ณผ๋ฅผ ๋‚˜ํƒ€๋‚ธ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ €ํฌ๊ฐ€ ์ œ์•ˆํ•œ ๊ธฐ๋ฒ•์ด ๋‹ค๋ฅธ Mesh Reconstruction ๋ฒ ์ด์Šค๋ผ์ธ ๊ธฐ๋ฒ•๋“ค์„ ์‚ฌ์šฉํ–ˆ์„ ๋•Œ ๋ณด๋‹ค ๋” ๋‚˜์€ Shape Deformation ๊ฒฐ๊ณผ๋ฅผ ๋ณด์ด๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

์œ„์˜ ๊ฒฐ๊ณผ์— ๋Œ€ํ•œ ์ •์„ฑ์  ๊ฒฐ๊ณผ (์•„๋ž˜ ๊ทธ๋ฆผ) ๋˜ํ•œ ์ €ํฌ ๊ธฐ๋ฒ•์ด ๋”์šฑ ์ž์—ฐ์Šค๋Ÿฌ์šด Shape Deformation์„ ๋ชจ๋ธ๋งํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.

3D-Aware Image Deformation

๋ณธ ๋™์˜์ƒ์€ ์ €ํฌ์˜ 3D-Aware Image Deformation ๊ธฐ๋ฒ•์„ ์ด์šฉํ•ด์„œ ์ƒ์„ฑํ•œ ๋ชจ์…˜ ๋™์˜์ƒ์ž…๋‹ˆ๋‹ค. Mesh Reconstruction ๋ฒ ์ด์Šค๋ผ์ธ ๊ธฐ๋ฒ•๋“ค๋ณด๋‹ค ๋”์šฑ ์ž์—ฐ์Šค๋Ÿฌ์šด Image Deformation์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.

Interactive Demo๋„ ์ฒดํ—˜ํ•ด๋ณด์‹œ๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค. ์‚ฌ์šฉ์ž๊ฐ€ ์ง๊ด€์ ์ธ Deformation Handle (Keypoint)๋ฅผ ์ด์šฉํ•˜์—ฌ ์˜์ƒ์„ ๋ณ€ํ˜•ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

5. Conclusion

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” Shape Laplacian์„ ํ•™์Šตํ•จ์œผ๋กœ์จ ๋ณด๋‹ค ์ž์—ฐ์Šค๋Ÿฌ์šด 3D-Aware Deformation์„ ๊ฐ€๋Šฅํ•˜๊ฒŒํ•˜๋Š” ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์•ˆํ•˜์˜€์Šต๋‹ˆ๋‹ค. ์ €ํฌ๊ฐ€ ์•Œ๊ธฐ๋กœ๋Š” ์ด๊ฐ€ ๋‰ด๋Ÿด๋„ท ๊ธฐ๋ฐ˜ ๊ธฐ๋ฒ•์ด Shape Lapacian ์˜ˆ์ธก์— ํšจ๊ณผ์ ์ผ ์ˆ˜ ์žˆ์Œ์„ ์ฒ˜์Œ์œผ๋กœ ๋ณด์ธ ์—ฐ๊ตฌ๋ผ๊ณ  ์•Œ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋ณธ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๋ฐœ์ „์‹œํ‚ค๊ธฐ ์œ„ํ•œ ๋”์šฑ ๋‹ค์–‘ํ•œ ์•„์ด๋””์–ด๊ฐ€ ๋งŽ์€๋ฐ, ๊ธฐํšŒ๊ฐ€ ๋œ๋‹ค๋ฉด ํ•ด๋‹น ๋ฐฉํ–ฅ์œผ๋กœ ๋”์šฑ ์—ฐ๊ตฌํ•ด๋ณด๊ณ  ์‹ถ์Šต๋‹ˆ๋‹ค.

Take-Home Message (์˜ค๋Š˜์˜ ๊ตํ›ˆ)

์ œ๊ฐ€ ๊ฐœ์ธ์ ์œผ๋กœ ์ด ํ”„๋กœ์ ํŠธ๋ฅผ ํ†ตํ•ด ๋ฐฐ์šด ๊ตํ›ˆ์€ "๋๊นŒ์ง€ ํฌ๊ธฐํ•˜์ง€ ์•Š๊ณ  ์ง‘๋…์„ ๊ฐ€์ง€๋ฉฐ ์—ฐ๊ตฌ ๋ฌธ์ œ๋ฅผ ํ’€์ž"๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ณธ ํ”„๋ ˆ์ž„์›Œํฌ ๊ฐœ๋ฐœ ๋‹จ๊ณ„์—์„œ ์ž์ž˜ํ•œ Challenge๋“ค์ด ๋งŽ์•˜์—ˆ๊ณ , ๊ทธ ๊ณผ์ • ์ค‘ ์›๋ž˜ ์ง„ํ–‰ํ•˜๋ ค๋˜ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์œผ๋กœ๋ถ€ํ„ฐ ํฌ๊ฒŒ ๋ฐ”๋€Œ์–ด ๋งˆ๋ฌด๋ฆฌ๋œ ๋ถ€๋ถ„๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ž˜๋„ ๋™์ž‘ํ•˜๋Š” ์†”๋ฃจ์…˜์„ ์ฐพ์•„๋‚ด๊ณ  ๋œป ๊นŠ๊ฒŒ ํ”„๋กœ์ ํŠธ๋ฅผ ๋งˆ๋ฌด๋ฆฌํ•  ์ˆ˜ ์žˆ์–ด์„œ ๊ฐœ์ธ์ ์œผ๋กœ๋Š” ๋งค์šฐ ๊ธฐ์–ต์— ๋‚จ๋Š” ์—ฐ๊ตฌ ๊ฒฝํ—˜์ด ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ณผ์ • ์ค‘ ํฐ ๋„์›€๊ณผ ์กฐ์–ธ์„ ์ฃผ์‹  ๋‘ ์ง€๋„ ๊ต์ˆ˜๋‹˜๊ป˜ ๊นŠ์€ ๊ฐ์‚ฌ๋ฅผ ๋“œ๋ฆฝ๋‹ˆ๋‹ค.

Author / Reviewer information

Author

์ด์ง€ํ˜„ (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 1031โ€“1040. SIAM, 2009.

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

Last updated

Was this helpful?