3 minute read

Reference

360-GS: Layout-guided Panoramic Gaussian Splatting For Indoor Roaming

[Equirectangular Image (๋“ฑ์žฅ๋ฐฉํ˜• ์ด๋ฏธ์ง€) ์„ค๋ช… ์ด๋ฏธ์ง€ ์ขŒํ‘œ ๋ณ€ํ™˜ ๊ตฌ๋ฉด์ขŒํ‘œ ๋ฒกํ„ฐ ๊ณ„์‚ฐ](https://mvje.tistory.com/211)

[๋…ผ๋ฌธ๋ฆฌ๋ทฐ] 360-GS: Layout-guided Panoramic Gaussian Splatting For Indoor Roaming

ํŒŒ๋…ธ๋ผ๋งˆ ์ด๋ฏธ์ง€๋Š” ๊ฐ€๋กœ๋กœ 360ยฐ ์ „์ฒด์™€ ์„ธ๋กœ๋กœ 180ยฐ ์ „์ฒด๋ฅผ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ์ ์œผ๋กœ, ์ด๋ฏธ์ง€์˜ ์ƒ๋‹จ๊ณผ ํ•˜๋‹จ์€ ์‹ฌ๊ฐํ•˜๊ฒŒ ์™œ๊ณก๋˜์–ด ๋‚˜ํƒ€๋‚ฉ๋‹ˆ๋‹ค.

Equirectangular image๋Š” ๊ตฌ๋ฉด ์ขŒํ‘œ๊ณ„๋ฅผ ์ง๊ฐ ์ขŒํ‘œ๊ณ„๋กœ ๋งคํ•‘ํ•œ ์ด๋ฏธ์ง€์ž…๋กœ ๊ตฌ๋ฉด ์ขŒํ‘œ๊ณ„์˜ ๊ฐ๋„๋ฅผ ์ง๊ฐ ์ขŒํ‘œ๊ณ„์˜ ํ”ฝ์…€๋กœ ์ผ๋Œ€์ผ ๋งคํ•‘ํ•˜์—ฌ ํ‘œํ˜„ํ•ฉ๋‹ˆ๋‹ค.

Equirectangular Projection(๋“ฑ ์žฅ๋ฐฉํ˜• ๋„๋ฒ•)์€ ๊ตฌ์ฒด ํ˜•ํƒœ์˜ ํŒŒ๋…ธ๋ผ๋งˆ ์ด๋ฏธ์ง€๋ฅผ ํ‘œ์‹œํ•˜๊ธฐ ์œ„ํ•ด ์ฃผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ํˆฌ์˜๋ฒ•์ž…๋‹ˆ๋‹ค.

๊ตฌ๋ฉด ์ขŒํ‘œ๊ณ„์—์„œ๋Š” ์œ„๋„(latitude)์™€ ๊ฒฝ๋„(longitude)๋กœ ์ด๋ฏธ์ง€๋ฅผ ํ‘œํ˜„ํ•˜๋Š”๋ฐ, ์œ„๋„๋Š” ๊ตฌ์˜ ์ˆ˜ํ‰ ๋ฐฉํ–ฅ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, -90์—์„œ 90๋„๊นŒ์ง€์˜ ๋ฒ”์œ„๋ฅผ ๊ฐ€์ง€๊ณ  ๊ฒฝ๋„๋Š” ๊ตฌ์˜ ์ˆ˜์ง ๋ฐฉํ–ฅ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, -180์—์„œ 180๋„๊นŒ์ง€์˜ ๋ฒ”์œ„๋ฅผ ๊ฐ€์ง‘๋‹ˆ๋‹ค.

Equirectangular ์ด๋ฏธ์ง€๋Š” ์ด๋Ÿฌํ•œ ๊ตฌ๋ฉด ์ขŒํ‘œ๊ณ„๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ด๋ฏธ์ง€๋ฅผ ํ‘œํ˜„ํ•ฉ๋‹ˆ๋‹ค.

์ด๋ฏธ์ง€์˜ ๋„ˆ๋น„๋Š” 360๋„์˜ ๊ฒฝ๋„ ๋ฒ”์œ„์— ๋งคํ•‘๋˜๊ณ , ๋†’์ด๋Š” -90์—์„œ 90๋„์˜ ์œ„๋„ ๋ฒ”์œ„์— ๋งคํ•‘๋ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด๋ฏธ์ง€์˜ ํฌ๊ธฐ๋Š” ๊ฒฝ๋„์™€ ์œ„๋„์˜ ํ•ด์ƒ๋„์— ๋”ฐ๋ผ ๊ฒฐ์ •๋ฉ๋‹ˆ๋‹ค.

์ด๋Ÿฌํ•œ Equirectangular ์ด๋ฏธ์ง€๋Š” ์ฃผ๋กœ 360๋„ ๋™์˜์ƒ, ๊ฐ€์ƒ ํ˜„์‹ค(VR) ์ด๋ฏธ์ง€, ํ–‰์„ฑ ํ‘œ๋ฉด ์ด๋ฏธ์ง€ ๋“ฑ๊ณผ ๊ฐ™์€ ๊ตฌํ˜• ๋ฐ์ดํ„ฐ๋ฅผ ํ‘œํ˜„ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š”๋ฐ, ์ด๋Ÿฌํ•œ ์ด๋ฏธ์ง€๋Š” ๊ตฌ๋ฉด ์ขŒํ‘œ๊ณ„์˜ ํŠน์„ฑ์„ ๋ณด์กดํ•˜๋ฉด์„œ๋„ ์ง๊ฐ ์ขŒํ‘œ๊ณ„์—์„œ์˜ ์ด๋ฏธ์ง€ ์ฒ˜๋ฆฌ ๋ฐ ๋ถ„์„์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค.

image


ํŒŒ๋…ธ๋ผ๋งˆ ํˆฌ์˜์—์„œ 3D-2D ๊ฐ€์šฐ์‹œ์•ˆ ๋งคํ•‘์˜ ํ•œ๊ณ„

Since the local affine approximation relies on projective transformation, it is not suitable for mapping 3D Gaussians to 2D Gaussians on panoramic images. A panoramic image covers the whole 360โ—ฆ horizontally and the whole 180โ—ฆ vertically.

Consequently, the top and bottom of the image appear severely distorted.

As illustrated in Fig. 3, the panoramic projection assumes distinct shapes that can not be modeled with Gaussians under varying configurations.

Employing a 2D Gaussian for fitting such a projection would lead to significant errors.

๋กœ์ปฌ ์•„ํ•€ ๊ทผ์‚ฌ๋Š” ํˆฌ์˜ ๋ณ€ํ™˜(projective transformation)์— ์˜์กดํ•˜๊ธฐ ๋•Œ๋ฌธ์—, 3D ๊ฐ€์šฐ์‹œ์•ˆ(3D Gaussians)์„ ํŒŒ๋…ธ๋ผ๋งˆ ์ด๋ฏธ์ง€(panoramic images) ์ƒ์˜ 2D ๊ฐ€์šฐ์‹œ์•ˆ(2D Gaussians)์œผ๋กœ ๋งคํ•‘ํ•˜๋Š” ๋ฐ ์ ํ•ฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.

ํŒŒ๋…ธ๋ผ๋งˆ ์ด๋ฏธ์ง€๋Š” ๊ฐ€๋กœ๋กœ 360ยฐ ์ „์ฒด์™€ ์„ธ๋กœ๋กœ 180ยฐ ์ „์ฒด๋ฅผ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ์ ์œผ๋กœ, ์ด๋ฏธ์ง€์˜ ์ƒ๋‹จ๊ณผ ํ•˜๋‹จ์€ ์‹ฌ๊ฐํ•˜๊ฒŒ ์™œ๊ณก๋˜์–ด ๋‚˜ํƒ€๋‚ฉ๋‹ˆ๋‹ค.

๊ทธ๋ฆผ 3์— ๋‚˜ํƒ€๋‚œ ๊ฒƒ์ฒ˜๋Ÿผ, ํŒŒ๋…ธ๋ผ๋งˆ ํˆฌ์˜(panoramic projection)์€ ๋‹ค์–‘ํ•œ ๊ตฌ์„ฑ(configuration) ํ•˜์—์„œ ๊ฐ€์šฐ์‹œ์•ˆ์œผ๋กœ ๋ชจ๋ธ๋งํ•  ์ˆ˜ ์—†๋Š” ๋…ํŠนํ•œ ํ˜•ํƒœ๋ฅผ ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค.

์ด๋Ÿฌํ•œ ํˆฌ์˜์— ๋Œ€ํ•ด 2D ๊ฐ€์šฐ์‹œ์•ˆ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ ํ•ฉ(fitting)ํ•˜๋ฉด ์ƒ๋‹นํ•œ ์˜ค๋ฅ˜๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

image

An alternative approach for the application of 3D-GS to panoramic inputs involves transforming the panoramas into perspective images before optimizing 3D Gaussians.

3DGS๋ฅผ panoramic inputs์— ์ ์šฉํ•˜๋Š” ๋Œ€์•ˆ์ ์ธ ์ ‘๊ทผ๋ฒ•์€, panormaic inputs์„ perspective image๋กœ ๋ณ€ํ™˜ํ•œ ํ›„ 3D ๊ฐ€์šฐ์‹œ์•ˆ์„ ์ตœ์ ํ™”ํ•˜๋Š” ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค.

๊ทธ๋Ÿฌ๋‚˜ ์ด์ฒ˜๋Ÿผ naiveํ•˜๊ฒŒ panoramic inputs์„ perspective image๋กœ ๋ณ€ํ™˜ํ›„ 3DGS์— ๋„ฃ์–ด 3DGS๋ฅผ ์ตœ์ ํ™”ํ•˜๊ณ ,

rendering๋œ perspective image๋“ค์„ ๋‹ค์‹œ panormaic inputs์œผ๋กœ stichingํ•˜๋ฉด stichiting artifacts๊ฐ€ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค.

๊ทธ๋ ‡๋‹ค๊ณ  ๋‹จ์ˆœํžˆ panormaic inputs์„ 3DGS์— ๋„ฃ์–ด ์ตœ์ ํ™”ํ•˜๋ฉด 3DGS๊ฐ€ ํŒŒ๋…ธ๋ผ๋งˆ ๋ฐ์ดํ„ฐ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์ฒ˜๋ฆฌํ•˜์ง€ ๋ชปํ•ด, ํ•™์Šต์ด ์‹คํŒจํ•ฉ๋‹ˆ๋‹ค.

image

  1. Left (์™ผ์ชฝ ํŒŒ์ดํ”„๋ผ์ธ)
  • ํŒŒ๋…ธ๋ผ๋งˆ ์ด๋ฏธ์ง€๋ฅผ ์—ฌ๋Ÿฌ ๊ฐœ์˜ ์›๊ทผ(perspective) ์ด๋ฏธ์ง€๋กœ ๋‚˜๋ˆ”:
    • ํŒŒ๋…ธ๋ผ๋งˆ ๋ฐ์ดํ„ฐ๋ฅผ ์ž‘์€ ์กฐ๊ฐ์œผ๋กœ ์ชผ๊ฐœ์„œ ์ผ๋ฐ˜์ ์ธ ์›๊ทผ ํˆฌ์˜ ์ด๋ฏธ์ง€๋ฅผ ๋งŒ๋“ญ๋‹ˆ๋‹ค.
    • ๊ฐ๊ฐ์˜ ์กฐ๊ฐ์€ ํŒŒ๋…ธ๋ผ๋งˆ ๋ฐ์ดํ„ฐ๋ฅผ ํŠน์ • ์นด๋ฉ”๋ผ ์œ„์น˜์—์„œ ๋ณธ ๊ฒƒ์ฒ˜๋Ÿผ ์ฒ˜๋ฆฌ๋ฉ๋‹ˆ๋‹ค.
  • 3D-GS์— ์ž…๋ ฅ:
    • ์ด๋ ‡๊ฒŒ ์ƒ์„ฑ๋œ ์›๊ทผ ์ด๋ฏธ์ง€๋ฅผ 3D-GS ๋ชจ๋ธ์— ์ž…๋ ฅ์œผ๋กœ ๋„ฃ์–ด ํ•™์Šต์„ ์ง„ํ–‰ํ•ฉ๋‹ˆ๋‹ค.
  • ํŒŒ๋…ธ๋ผ๋งˆ ์žฌ๊ตฌ์„ฑ:
    • 3D-GS๋ฅผ ์‚ฌ์šฉํ•ด ์›๊ทผ ์ด๋ฏธ์ง€๋ฅผ ๋ Œ๋”๋งํ•œ ํ›„, ์ด๋ฅผ ๋‹ค์‹œ ํŒŒ๋…ธ๋ผ๋งˆ๋กœ ํ•ฉ์นฉ๋‹ˆ๋‹ค.
    • ๋ฌธ์ œ์ : ์ด ๊ณผ์ •์—์„œ ์ด๋ฏธ์ง€ ์กฐ๊ฐ๋“ค์„ ์ด์–ด๋ถ™์ด๋Š”(stitching) ๊ณผ์ •์—์„œ ์—ฐ๊ฒฐ ๋ถ€์œ„์— ์ธ๊ณต์ ์ธ ์™œ๊ณก(stitching artifacts)์ด ์ƒ๊น๋‹ˆ๋‹ค.
  1. Right (์˜ค๋ฅธ์ชฝ ํŒŒ์ดํ”„๋ผ์ธ)
  • ํŒŒ๋…ธ๋ผ๋งˆ ๋ฐ์ดํ„ฐ๋ฅผ ๊ทธ๋Œ€๋กœ ์ž…๋ ฅ:
    • ํŒŒ๋…ธ๋ผ๋งˆ ๋ฐ์ดํ„ฐ๋ฅผ ์›๊ทผ ์ด๋ฏธ์ง€๋กœ ์ชผ๊ฐœ์ง€ ์•Š๊ณ , ๊ทธ๋Œ€๋กœ 3D-GS์— ์ง์ ‘ ์ž…๋ ฅํ•ฉ๋‹ˆ๋‹ค.
  • ๋ฌธ์ œ์ :
    • 3D-GS๋Š” ๊ธฐ๋ณธ์ ์œผ๋กœ ์›๊ทผ ํˆฌ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์ฒ˜๋ฆฌํ•˜๋„๋ก ์„ค๊ณ„๋˜์–ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์—, ํŒŒ๋…ธ๋ผ๋งˆ ์ž…๋ ฅ์— ์ ํ•ฉํ•˜์ง€ ์•Š์•„์„œ ํ•™์Šต์ด ์‹คํŒจํ•ฉ๋‹ˆ๋‹ค.

์š”์•ฝ

  • Left ๋ฐฉ๋ฒ•: ํŒŒ๋…ธ๋ผ๋งˆ๋ฅผ ์ž‘์€ ์›๊ทผ ์ด๋ฏธ์ง€๋กœ ๋‚˜๋ˆˆ ๋’ค ํ•™์Šตํ•˜๊ณ , ์ด๋ฅผ ๋‹ค์‹œ ํ•ฉ์น˜๋Š” ๋ฐฉ์‹ โ†’ โ€œ์ด์–ด๋ถ™์ด๋Š” ๊ณผ์ •์—์„œ ์™œ๊ณก ๋ฐœ์ƒโ€
  • Right ๋ฐฉ๋ฒ•: ํŒŒ๋…ธ๋ผ๋งˆ๋ฅผ ๊ทธ๋Œ€๋กœ ์ž…๋ ฅ์œผ๋กœ ์‚ฌ์šฉ โ†’ โ€œ3DGS๊ฐ€ ํŒŒ๋…ธ๋ผ๋งˆ ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๋Œ€๋กœ ์ฒ˜๋ฆฌํ•˜์ง€ ๋ชปํ•ด ํ•™์Šต ์‹คํŒจโ€
  • ๋”ฐ๋ผ์„œ, ๋‘ ๋ฐฉ๋ฒ• ๋ชจ๋‘ ํŒŒ๋…ธ๋ผ๋งˆ ๋ฐ์ดํ„ฐ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์ฒ˜๋ฆฌํ•˜๊ธฐ์—๋Š” ๋ฌธ์ œ๊ฐ€ ์žˆ๋‹ค๋Š” ์ ์„ ์ง€์ ํ•œ ๋‚ด์šฉ์ž…๋‹ˆ๋‹ค.

360-GS์˜ ์ €์ž๋Š” ์œ„์™€ ๊ฐ™์ด panormaic inputs์„ ๊ทธ๋Œ€๋กœ ์“ฐ๋Š” ๊ฒƒ์— ๋ฌธ์ œ๊ฐ€ ์žˆ์Œ์„ ์ฐฉ์•ˆํ•˜๊ณ , ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด tangent plane์— ๋จผ์ € 3d gaussian์„ splattingํ•˜๊ณ  spherical surface๋กœ mappingํ•˜๋Š” ๋ฐฉ์‹์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.

๋ชฉํ‘œ๋Š” ํŒŒ๋…ธ๋ผ๋งˆ ์ง‘ํ•ฉ์—์„œ 3D ๊ฐ€์šฐ์‹œ์•ˆ ํ‘œํ˜„์„ ์ตœ์ ํ™”ํ•˜๊ณ , ์ง์ ‘์ ์ธ ํŒŒ๋…ธ๋ผ๋งˆ ๋ Œ๋”๋ง์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.

๊ตฌ๋ฉด ํˆฌ์˜์„ ์ง์ ‘์ ์œผ๋กœ ํ‘œํ˜„ํ•˜๋Š” ๋ฐ ์žˆ์–ด ๋ฐœ์ƒํ•˜๋Š” ์–ด๋ ค์›€์„ ๊ณ ๋ คํ•˜์—ฌ, ์ €์ž๋Š” ๊ตฌ๋ฉด ํ‘œ๋ฉด(spherical surface)์— ๋Œ€ํ•œ splatting์„ ๋‘ ๋‹จ๊ณ„๋กœ ๋ถ„ํ•ดํ•˜๋Š” ์ƒˆ๋กœ์šด splatting ๊ธฐ๋ฒ•์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค.

  • ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„๋Š” ๋‹จ์œ„ ๊ตฌ์˜ ์ ‘์„  ํ‰๋ฉด(tangent plane)์—์„œ splatting์„ ์ˆ˜ํ–‰ํ•˜๊ณ ,
  • ๋‘ ๋ฒˆ์งธ ๋‹จ๊ณ„๋Š” ์ด๋ฅผ ๊ตฌ๋ฉด ํ‘œ๋ฉด(spherical surface)์œผ๋กœ ๋งคํ•‘ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.

์ด๋ฅผ ํ†ตํ•ด 3D ๊ฐ€์šฐ์‹œ์•ˆ์„ 2D ๊ฐ€์šฐ์‹œ์•ˆ์œผ๋กœ ํˆฌ์˜ํ•˜์—ฌ ๋ Œ๋”๋ง์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค.

360ยฐ ๊ฐ€์šฐ์‹œ์•ˆ splatting ๊ธฐ๋ฒ•์˜ ๊ฐœ์š”๋Š” ๊ทธ๋ฆผ 5์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค.

image

Leave a comment