๋ณธ๋ฌธ ๋ฐ”๋กœ๊ฐ€๊ธฐ
๐Ÿ˜ŽAI/Terminology

[Notable] NeuS (Neural Implicit Surface)

by SolaKim 2025. 3. 18.

https://lingjie0206.github.io/papers/NeuS/

 

NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-view Reconstruction

Abstract We present a novel neural surface reconstruction method, called NeuS (pronunciation: /nuหz/, same as "news"), for reconstructing objects and scenes with high fidelity from 2D image inputs. Existing neural surface reconstruction approaches, such a

lingjie0206.github.io

 

 

 

NeuS(Neural Implicit Surface) ๋Š” ์‹ ๊ฒฝ ๊ธฐ๋ฐ˜ ์•”์‹œ์  ํ‘œ๋ฉด ํ•™์Šต ๊ธฐ๋ฒ•์œผ๋กœ, NeRF(Neural Radiance Fields) ๊ธฐ๋ฐ˜์˜ 3D ์žฌ๊ตฌ์„ฑ ๊ธฐ์ˆ  ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. 

๊ธฐ์กด NeRF ์™€ ๋‹ฌ๋ฆฌ, NeuS ๋Š” ๋ช…ํ™•ํ•œ ํ‘œ๋ฉด(Surface) ์„ ์ถ”์ •ํ•˜๋Š”๋ฐ ์ดˆ์ ์„ ๋งž์ถ”๋ฉฐ, ๋ฌผ์ฒด์˜ ์ •ํ™•ํ•œ ํ˜•์ƒ์„ ์ถ”์ถœํ•˜๊ณ  ๋ Œ๋”๋งํ•˜๋Š” ๋ฐ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค.

 

1. ๊ธฐ์กด NeRF ์™€์˜ ์ฐจ์ด์ 

NeRF ๋Š” 3D ๊ณต๊ฐ„์„ ๋ฐ€๋„(density)๋กœ ํ‘œํ˜„ํ•˜๋Š” ๋ณผ๋ฅจ ๋ Œ๋”๋ง(Volume Rendering) ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•˜์ง€๋งŒ, ์ด๋Š” ๋ช…ํ™•ํ•œ ํ‘œ๋ฉด(surface) ์„ ์ถ”์ถœํ•˜๋Š” ๋ฐ ํ•œ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.

NeRF์˜ ๋ฌธ์ œ์ 

  • 3D ๊ณต๊ฐ„์—์„œ ๋ฐ€๋„(σ)๋ฅผ ์ง์ ‘ ํ•™์Šต → ํŠน์ • ํ‘œ๋ฉด์„ ์ •ํ™•ํžˆ ์ •์˜ํ•˜๊ธฐ ์–ด๋ ค์›€.
  • ํ‘œ๋ฉด์ด ์•„๋‹Œ, ๋ฌผ์ฒด์˜ ๋‚ด๋ถ€ ๊ตฌ์กฐ๊นŒ์ง€ ํ•จ๊ป˜ ํ‘œํ˜„๋จ.
  • ๋…ธ์ด์ฆˆ๊ฐ€ ๋งŽ๊ณ  ๋šœ๋ ทํ•œ ๊ฒฝ๊ณ„๋ฅผ ์œ ์ง€ํ•˜๊ธฐ ์–ด๋ ค์›€.

NeuS์˜ ๊ฐœ์„ ์ 

  • ๋ฐ€๋„๊ฐ€ ์•„๋‹Œ Signed Distance Function (SDF) ์„ ํ•™์Šตํ•˜์—ฌ ๋ช…ํ™•ํ•œ ํ‘œ๋ฉด์„ ์ถ”์ •
  • SDF๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ‘œ๋ฉด ์œ„์น˜๋ฅผ ์ง์ ‘ ๊ฒฐ์ •ํ•˜๋ฏ€๋กœ ๊ฒฝ๊ณ„๊ฐ€ ๋šœ๋ ทํ•˜๊ณ  ๋งค๋„๋Ÿฌ์šด 3D ํ˜•์ƒ์„ ์ƒ์„ฑ
  • ๋”์šฑ ์ •ํ™•ํ•œ 3D ์žฌ๊ตฌ์„ฑ์ด ๊ฐ€๋Šฅํ•˜๋ฉฐ, ๋ณต์žกํ•œ ๋ฌผ์ฒด์—์„œ๋„ ๊นจ๋—ํ•œ ๋ฉ”์‰ฌ๋ฅผ ์ƒ์„ฑ

 

2. NeuS ์˜ ํ•ต์‹ฌ ๊ฐœ๋…

NeuS ๋Š” SDF(Signed Distance Function) ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ํ‘œ๋ฉด์„ ์ถ”์ •ํ•˜๊ณ , ๋ณผ๋ฅจ ๋ Œ๋”๋ง ๋ฐฉ์‹์„ ์ˆ˜์ •ํ•˜์—ฌ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์ด ํŠน์ง•์ž…๋‹ˆ๋‹ค.

2.1 Signed Distance Function (SDF)

  • SDF ๋Š” ๊ณต๊ฐ„์˜ ํ•œ ์  x ๊ฐ€ ๋ฌผ์ฒด ํ‘œ๋ฉด์œผ๋กœ๋ถ€ํ„ฐ ์–ผ๋งˆ๋‚˜ ๋–จ์–ด์ ธ ์žˆ๋Š”์ง€๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ํ•จ์ˆ˜
  • ํ‘œ๋ฉด์— ๊ฐ€๊นŒ์šธ์ˆ˜๋ก 0์— ๊ฐ€๊นŒ์›Œ์ง€๊ณ , ๋ฌผ์ฒด ๋‚ด๋ถ€๋Š” ์Œ์ˆ˜, ์™ธ๋ถ€๋Š” ์–‘์ˆ˜ ๊ฐ’์„ ๊ฐ€์ง
  • NeuS ๋Š” ์ด SDF ๋ฅผ ์‹ ๊ฒฝ๋ง(MLP) ์œผ๋กœ ํ•™์Šตํ•˜์—ฌ ๋ฌผ์ฒด ํ‘œ๋ฉด์„ ์ง์ ‘ ์ฐพ์•„๋ƒ„

 

2.2 NeuS ์˜ ์ƒˆ๋กœ์šด ๋ณผ๋ฅจ ๋ Œ๋”๋ง ๊ณต์‹

๊ธฐ์กด NeRF ๋Š” ๋ฐ€๋„๋ฅผ ์‚ฌ์šฉํ•œ ๋ณผ๋ฅจ ๋ Œ๋”๋ง์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ–ˆ์ง€๋งŒ, NeuS ๋Š” SDF ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ‘œ๋ฉด์„ ์ถ”์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด NeuS๋Š” SDF ๊ฐ’์„ ํ™•๋ฅ  ๋ฐ€๋„ ํ•จ์ˆ˜ (PDF) ๋กœ ๋ณ€ํ™˜ํ•˜์—ฌ ํ‘œ๋ฉด์„ ์ƒ˜ํ”Œ๋งํ•ฉ๋‹ˆ๋‹ค.

  • SDF ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ‘œ๋ฉด์„ ์ฐพ๊ณ , ํ•ด๋‹น ์œ„์น˜์—์„œ ์ƒ‰์ƒ์„ ์ƒ˜ํ”Œ๋ง
  • ๋ฌผ์ฒด ํ‘œ๋ฉด ๊ทผ์ฒ˜์—์„œ๋งŒ ์ƒ˜ํ”Œ๋งํ•˜๋ฏ€๋กœ ๋”์šฑ ํšจ์œจ์ ์ธ ํ•™์Šต ๊ฐ€๋Šฅ
  • ๊ธฐ์กด NeRF ๋ณด๋‹ค ์ ์€ ์ƒ˜ํ”Œ๋กœ๋„ ์ •๋ฐ€ํ•œ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Œ

 

3. NeuS ์˜ ํ•™์Šต ๊ณผ์ •

1๏ธโƒฃ Signed Distance Function (SDF) ํ•™์Šต

  • MLP๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ ์ ์˜ SDF ๊ฐ’์„ ์˜ˆ์ธก
  • ํ‘œ๋ฉด์„ ๋‚˜ํƒ€๋‚ด๋Š” 0 ๊ฐ’ ์ฃผ๋ณ€์„ ์ •ํ™•ํžˆ ๋ชจ๋ธ๋งํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”

2๏ธโƒฃ SDF → ๋ฐ€๋„(α) ๋ณ€ํ™˜

  • SDF ๊ฐ’์„ ํ™•๋ฅ  ๋ฐ€๋„ ํ•จ์ˆ˜(PDF) ๋กœ ๋ณ€ํ™˜ํ•˜์—ฌ ๋ณผ๋ฅจ ๋ Œ๋”๋ง์— ํ™œ์šฉ
  • ์ด๋ฅผ ํ†ตํ•ด NeRF์ฒ˜๋Ÿผ ๋ณผ๋ฅจ ๋ Œ๋”๋ง์ด ๊ฐ€๋Šฅํ•˜์ง€๋งŒ, ํ‘œ๋ฉด์„ ๋ช…ํ™•ํ•˜๊ฒŒ ํ‘œํ˜„

3๏ธโƒฃ ๋ Œ๋”๋ง ๋ฐ ์ƒ‰์ƒ ์˜ˆ์ธก

  • ์นด๋ฉ”๋ผ ๋ฐฉํ–ฅ์— ๋”ฐ๋ฅธ ๊ด‘์„ (ray)์„ ์ถ”์ ํ•˜์—ฌ ํ‘œ๋ฉด์— ํ•ด๋‹นํ•˜๋Š” ์ƒ‰์ƒ์„ ์˜ˆ์ธก
  • ํ‘œ๋ฉด ๋ฒ•์„ (normals) ์ •๋ณด๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๋”์šฑ ์‚ฌ์‹ค์ ์ธ ์กฐ๋ช… ํšจ๊ณผ ์ ์šฉ ๊ฐ€๋Šฅ

4๏ธโƒฃ ์ตœ์ ํ™”

  • ๋ Œ๋”๋ง๋œ ์ด๋ฏธ์ง€์™€ ์‹ค์ œ ์ด๋ฏธ์ง€ ์‚ฌ์ด์˜ L2 loss, SDF regularization loss ๋“ฑ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ ํ•™์Šต