2 minute read

https://github.com/Anttwo/SuGaR/issues/2

Equation (4)

\[(p - \mu_g)^T \Sigma_g^{-1} (p - \mu_g) \approx \frac{1}{s_g^2} \langle p - \mu_g, n_g \rangle^2\]
  • Gaussians๊ฐ€ โ€˜flatโ€™ํ•˜๋‹ค๋Š” ๊ฒƒ์€ ํ•˜๋‚˜์˜ scaling factor๊ฐ€ ๋‹ค๋ฅธ ๊ฒƒ๋“ค๋ณด๋‹ค ํ›จ์”ฌ ์ž‘๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.
  • inverse covariance matrix๋ฅผ ๋Œ€๊ฐ์„ ์˜ ํ•˜๋‚˜์˜ non-zero ๊ฐ’๊ณผ ๋‚˜๋จธ์ง€ ๋‘ ๊ฐœ์˜ ์ž‘์€ ๊ฐ’๋งŒ ์žˆ๋Š” ๋Œ€๊ฐ ํ–‰๋ ฌ๋กœ ๊ทผ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
  • ์ด non-zero ๊ฐ’์ด ํˆฌ์˜๋œ ๊ฒฐ๊ณผ๋ฅผ ์Šค์ผ€์ผ๋งํ•ฉ๋‹ˆ๋‹ค.

์งˆ๋ฌธ: ์™œ Equation (4)๊ฐ€ ์„ฑ๋ฆฝํ•˜๋Š”๊ฐ€?

๋‹ต๋ณ€: Equation (4)๋Š” Gaussians๊ฐ€ โ€˜flatโ€™ํ•œ ์ด์ƒ์ ์ธ ๊ฒฝ์šฐ์˜ ๊ทผ์‚ฌ์—์„œ ๋‚˜์˜ต๋‹ˆ๋‹ค. โ€˜Flatโ€™ํ•œ Gaussian์€ ํ•˜๋‚˜์˜ scaling factor๊ฐ€ ๋‹ค๋ฅธ ๊ฒƒ๋“ค๋ณด๋‹ค ํ›จ์”ฌ ์ž‘์€ ๊ฒฝ์šฐ๋ฅผ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ์ด ์ž‘์€ scaling factor๋ฅผ $\lambda$๋ผ๊ณ  ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.

Rotation matrix $R$์™€ diagonal scaling matrix $S$์˜ ์—ญํ–‰๋ ฌ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ „์ฒด inverse covariance matrix $\Sigma^{-1}$๋ฅผ ์ž‘์„ฑํ•˜๋ฉด, $\Sigma^{-1}$๋ฅผ ๋Œ€๊ฐ์„ ์˜ ํ•˜๋‚˜์˜ non-zero ๊ฐ’๊ณผ ๋‚˜๋จธ์ง€ ๋‘ ๊ฐœ์˜ ์ž‘์€ ๊ฐ’๋งŒ ์žˆ๋Š” ๋Œ€๊ฐ ํ–‰๋ ฌ๋กœ ๊ทผ์‚ฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด non-zero ๊ฐ’์€ $\lambda$์ž…๋‹ˆ๋‹ค.

๋”ฐ๋ผ์„œ $\Sigma^{-1}$๋ฅผ $\alpha$๋กœ ๊ณฑํ•˜๋Š” ๊ฒƒ์€ $\alpha$๋ฅผ $\Sigma^{-1}$์˜ ๊ฐ€์žฅ ์ž‘์€ scaling factor์™€ ์—ฐ๊ด€๋œ ์—ด์— ํˆฌ์˜ํ•˜๊ณ  ๊ฒฐ๊ณผ๋ฅผ $\lambda$๋กœ ์Šค์ผ€์ผ๋งํ•˜๋Š” ๊ฒƒ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค. ์ด ์—ด์ด $\alpha$์™€ ์—ฐ๊ด€๋œ scaling axis์— ํ•ด๋‹นํ•˜๊ธฐ ๋•Œ๋ฌธ์— Equation (4)๊ฐ€ ์„ฑ๋ฆฝํ•ฉ๋‹ˆ๋‹ค.

Equation (5)

image

Equation (6)

\[\overline{f}(p) = \pm s_{g^*} \sqrt{-2 \log(\overline{d}(p))}\]
  • SDF๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ •๊ทœํ™”ํ•˜๋ฉด ๋” ๋‚˜์€ ์žฌ๊ตฌ์„ฑ ๊ฒฐ๊ณผ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
  • density ํ•จ์ˆ˜์—์„œ ๊ฑฐ๋ฆฌ ํ•จ์ˆ˜๋กœ ๋ณ€ํ™˜ํ•˜๋ฉด ํšจ๊ณผ์ ์œผ๋กœ ์žฌ๊ตฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
  • Equation (6)์€ Equation (5)์˜ ์—ญ๊ณต์‹์ž…๋‹ˆ๋‹ค.

์งˆ๋ฌธ: Equation (6)์—์„œ opacity ๋Œ€์‹  SDF๋ฅผ ์ •๊ทœํ™”ํ•˜๋Š” ์ด์œ ๋Š” ๋ฌด์—‡์ธ๊ฐ€? ์ด ํŠน์ •ํ•œ ๊ณต์‹์„ ์–ด๋–ป๊ฒŒ ๋„์ถœํ–ˆ๋Š”๊ฐ€?

๋‹ต๋ณ€: density ํ•จ์ˆ˜๋ฅผ Signed Distance Function (SDF)์œผ๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ๊ฒƒ์ด ์žฌ๊ตฌ์„ฑ์„ ๋” ์ž˜ ์ •๊ทœํ™”ํ•ฉ๋‹ˆ๋‹ค. SDF๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ์ •๋Ÿ‰์ ์ธ ๊ฒฐ๊ณผ๊ฐ€ ๋” ์ข‹์•„์ง€๊ณ  ๋ฐฐ๊ฒฝ๋„ ๋” ์ž˜ ์ •๊ทœํ™”๋ฉ๋‹ˆ๋‹ค.

์ง๊ด€์ ์œผ๋กœ ๋ณด๋ฉด, Equation (5)๋Š” ํ‰ํ‰ํ•˜๊ฒŒ ๋ถ„ํฌ๋œ Gaussians์˜ density ํ•จ์ˆ˜๋ฅผ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ์ง€์ˆ˜ ๋‚ด๋ถ€์˜ ์Šค์นผ๋ผ ๊ณฑ์€ 3D ์  $p$์™€ 3D Gaussian์˜ ์ค‘์‹ฌ์„ ํ†ต๊ณผํ•˜๋Š” ํ‰๋ฉด ์‚ฌ์ด์˜ ๊ฑฐ๋ฆฌ์™€ ๊ฐ™์Šต๋‹ˆ๋‹ค. ๋ฒกํ„ฐ $v$๋Š” Gaussian์˜ ๊ฐ€์žฅ ์ž‘์€ scaling factor์™€ ์—ฐ๊ด€๋œ ์ถ•์ด๋ฏ€๋กœ, Gaussian์ด ํ‰ํ‰ํ•  ๋•Œ ๊ฐ€์žฅ ์ž‘์€ scaling factor๊ฐ€ 0์— ๊ฐ€๊นŒ์›Œ์ง€๊ณ , ์ด ๋ฒกํ„ฐ๋ฅผ ํ‘œ๋ฉด์˜ normal๋กœ ๊ฐ„์ฃผํ•˜๋Š” ๊ฒƒ์ด ๋งค์šฐ ์ง๊ด€์ ์ž…๋‹ˆ๋‹ค.

๋งˆ์ง€๋ง‰์œผ๋กœ, Equation (6)์€ ๋‹จ์ˆœํžˆ Equation (5)์˜ ์—ญ๊ณต์‹์ž…๋‹ˆ๋‹ค. ๋งŒ์•ฝ Equation (5)๊ฐ€ $d(p) = h(\langle p - \mu_g, n_g \rangle)$์™€ ๊ฐ™์€ ๊ฐ’์„ ์ œ๊ณตํ•œ๋‹ค๋ฉด, Equation (6)์€ $\langle p - \mu_g, n_g \rangle = h^{-1}(d(p))$๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ด€๊ณ„๋Š” ์ด์ƒ์ ์ธ ๊ฒฝ์šฐ์— ์„ฑ๋ฆฝํ•˜์ง€๋งŒ, ์‹ค์ œ ๋น„์ด์ƒ์ ์ธ density $\rho$๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ Equation (6)์— ๋”ฐ๋ผ $\rho$๋ฅผ ๊ณ„์‚ฐํ•˜์—ฌ density๊ฐ€ non-destructive wayํ•˜๊ฒŒ ์ด์ƒ์ ์ธ density ํ•จ์ˆ˜๋กœ ์ˆ˜๋ ดํ•˜๋„๋ก ๊ฐ•์ œํ•ฉ๋‹ˆ๋‹ค.

Fig.5

image

  • depth map ๊ทผ์‚ฌ์น˜๋Š” SDF๋ฅผ ๊ณ„์‚ฐํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค.
  • ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ํ‘œ๋ฉด ์  ๊ทผ์‚ฌ์น˜๋Š” depth map์˜ ํˆฌ์˜์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
  • ์ •๊ทœํ™”๋Š” Gaussians๊ฐ€ ํ‘œ๋ฉด๊ณผ ์ผ์น˜ํ•˜๊ณ  ์นด๋ฉ”๋ผ ํฌ์ฆˆ๋ฅผ ํ–ฅํ•˜๊ฒŒ ํ•˜์—ฌ ๋ฐฐ๊ฒฝ์„ ์ •๊ทœํ™”ํ•ฉ๋‹ˆ๋‹ค.

์งˆ๋ฌธ: ์™œ Fig. 5์—์„œ๋Š” ์  $p$์˜ ๊นŠ์ด์˜ ์ฐจ์ด๋ฅผ ๊ณ„์‚ฐํ•˜์—ฌ $f(p)$๋ฅผ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ๋Š”๊ฐ€?

๋‹ต๋ณ€: estimator $\hat{f}(p)$๋Š” ํ˜„์žฌ ์žฅ๋ฉด๊ณผ ์—ฐ๊ด€๋œ ์‹ค์ œ SDF์˜ ๋Œ€๋žต์ ์ธ ๊ทผ์‚ฌ์น˜์ž…๋‹ˆ๋‹ค. ์ด ๊ทผ์‚ฌ๋Š” โ€œsplattedโ€ depth maps์˜ ๋งฅ๋ฝ์—์„œ ์˜๋ฏธ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.

์ตœ์ ํ™” ์ค‘์ธ ์นด๋ฉ”๋ผ $c$์— ๋Œ€ํ•ด Gaussian Splatting rasterizer๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ depth map์„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ๋น„๋ก ์™„๋ฒฝํ•˜์ง€๋Š” ์•Š์ง€๋งŒ, ์ด depth map์€ ์žฅ๋ฉด์˜ ํ‘œ๋ฉด์„ ์ž˜ ๋ฌ˜์‚ฌํ•ฉ๋‹ˆ๋‹ค. Gaussian ๋ถ„ํฌ์˜ ๊ณฑ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ˜ํ”Œ๋ง๋œ ์  $p$๋Š” ์‹ค์ œ ํ‘œ๋ฉด ๊ทผ์ฒ˜์— ์œ„์น˜ํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์œผ๋ฉฐ, $p$์˜ SDF ๊ฐ’, ์ฆ‰ ํ‘œ๋ฉด๊นŒ์ง€์˜ ๊ฑฐ๋ฆฌ๋Š” $p$์™€ ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด depth map์˜ ํ‘œ๋ฉด ์  $q$ ์‚ฌ์ด์˜ ๊ฑฐ๋ฆฌ์™€ ๊ฐ™์Šต๋‹ˆ๋‹ค.

depth map์ด โ€œsplattedโ€๋˜์–ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๊ด€์ฐฐ๋œ ํ‘œ๋ฉด์€ ์นด๋ฉ”๋ผ๋ฅผ ํ–ฅํ•œ ์ž‘์€ ์š”์†Œ๋“ค๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ $p$์™€ ๋™์ผํ•œ ์‹œ์„  ์ƒ์— ์žˆ๋Š” ์ ์ด depth map์˜ ํˆฌ์˜์— ํ•ด๋‹นํ•˜๋Š” ์  $q$์ผ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์Šต๋‹ˆ๋‹ค.

์ด ์ ‘๊ทผ๋ฒ•์€ ๋ฐ€๋„์— ๋Œ€ํ•œ ๊นŠ์ด ์ •๊ทœํ™”๋ฅผ ํฌํ•จ์‹œํ‚ค๊ธฐ ์œ„ํ•œ ๋„๊ตฌ๋กœ, Gaussians๊ฐ€ ํ‘œ๋ฉด๊ณผ ์ผ์น˜ํ•˜๋„๋ก ์œ ๋„ํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ, ์นด๋ฉ”๋ผ ํฌ์ฆˆ๋ฅผ ํ‰๊ท ์ ์œผ๋กœ ํ–ฅํ•˜๊ฒŒ ํ•˜๋Š” ์œ ์šฉํ•œ ์šฐ์„  ์ˆœ์œ„๋ฅผ ์ œ๊ณตํ•˜์—ฌ ๋ฐฐ๊ฒฝ์„ ์ •๊ทœํ™”ํ•˜๋Š” ๋ฐ ๋„์›€์„ ์ค๋‹ˆ๋‹ค.

Leave a comment