5 minute read

COLMAP์˜ Structure from Motion (SfM)

SfM is the process of reconstructing 3D structure from its projections into a series of images taken from different viewpoints.

image

COLMAP์—์„œ๋Š” ์•„๋ž˜์™€ ๊ฐ™์€ ์ฃผ์˜์‚ฌํ•ญ์ด ์žˆ์–ด์„œ ๋„ค๊ฐ€์ง€ ์กฐ๊ฑด์ด ์ž˜ ์ถฉ์กฑ๋˜์ง€ ์•Š์œผ๋ฉด 3D reconstruction ๊ฒฐ๊ณผ๊ฐ€ ์ข‹์ง€ ์•Š์Šต๋‹ˆ๋‹ค.

ํŠนํžˆ texture ์ •๋ณด๊ฐ€ ๋งŽ์ด ์—†๋Š”(ํŒจํ„ด์ด ๊ฑฐ์˜ ์—†๋Š”) ๊ฐ์ฒด๋Š” ํ˜•์ƒ์ด ๋งŽ์ด ๋ง๊ฐ€์ง„์ฑ„๋กœ ๋ณต์›๋ฉ๋‹ˆ๋‹ค.

  • Texture ๊ฐ€ ์ข‹์€ ์ด๋ฏธ์ง€ ์‚ฌ์šฉ
  • ์œ ์‚ฌํ•œ ์กฐ๋ช… ์กฐ๊ฑด์˜ ์ด๋ฏธ์ง€ ์‚ฌ์šฉ
  • ์‹œ๊ฐ์ ์œผ๋กœ ๋งŽ์ด ์ค‘์ฒฉ๋œ ์ด๋ฏธ์ง€ ์‚ฌ์šฉ
  • ๋‹ค์–‘ํ•œ viewpoints์—์„œ ๊ด€์ธกํ•œ ์ด๋ฏธ์ง€ ์‚ฌ์šฉ

SfM์€ ๋จผ์ € ๋‹ค์Œ 3๊ฐ€์ง€๋ฅผ ์ˆœ์„œ๋Œ€๋กœ ์ˆ˜ํ–‰ํ•˜๊ณ , scene graph๋ฅผ Incremental Reconstruction์— ๋„˜๊ฒจ์ค๋‹ˆ๋‹ค.

Correspondence Search

  1. feature extraction
  2. matching
  3. geometric verification

Incremental Reconstruction

reconstruction stage์—์„œ

  • input: scene graph
  • outputs:
    • pose estimates: $\mathcal{P}=$ { $\mathbf{P}_c \in \mathbf{SE}(3) \mid c=1โ€ฆN_P$ } for registered images.
    • the reconstructed scene structure as a set of points $\mathcal{X}=$ { $\mathbf{X}_k \in \mathbb{R}^3 \mid k=1โ€ฆN_X$ }.

Incremental Reconstruction์˜ workflow๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.

Correspondence Search์˜ ๊ฒฐ๊ณผ๋ฌผ์ธ Scene Graph๋Š” ์žฌ๊ตฌ์„ฑ ๋‹จ๊ณ„์˜ ๊ธฐ์ดˆ๊ฐ€ ๋˜๋ฉฐ, ๋ชจ๋ธ์„ ์‹ ์ค‘ํ•˜๊ฒŒ ์„ ํƒ๋œ ๋‘ ๊ฐœ์˜ ๋ทฐ ์žฌ๊ตฌ์„ฑ์œผ๋กœ ์ดˆ๊ธฐํ™”ํ•œ ํ›„, ์ ์ง„์ ์œผ๋กœ ์ƒˆ๋กœ์šด ์ด๋ฏธ์ง€๋ฅผ ๋“ฑ๋กํ•˜๊ณ , ์”ฌ ํฌ์ธํŠธ๋ฅผ ์‚ผ๊ฐ์ธก๋Ÿ‰ํ•˜๊ณ , ์•„์›ƒ๋ผ์ด์–ด๋ฅผ ํ•„ํ„ฐ๋งํ•˜๋ฉฐ, Bundle Adjustment (BA)์„ ์‚ฌ์šฉํ•˜์—ฌ ์žฌ๊ตฌ์„ฑ์„ refineํ•ฉ๋‹ˆ๋‹ค.

Incremental Reconstruction์˜ ๋‹จ๊ณ„๋ฅผ ์„ค๋ช…ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.

Initialization, Image Registration, Triangulation, Bundle Adjustment

1. Initialization: ๋ชจ๋ธ์„ ์‹ ์ค‘ํ•˜๊ฒŒ ์„ ํƒ๋œ ๋‘ ๊ฐœ์˜ ๋ทฐ(์ด๋ฏธ์ง€)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ดˆ๊ธฐํ™”ํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ โ€œ์ดˆ๊ธฐํ™”โ€๋ž€, ๋ชจ๋ธ์˜ ์‹œ์ž‘์ ์„ ์„ค์ •ํ•˜๋Š” ๊ณผ์ •์œผ๋กœ, ์„ ํƒ๋œ ๋‘ ๊ฐœ์˜ ์ด๋ฏธ์ง€๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ดˆ๊ธฐ 3D ๊ตฌ์กฐ๋ฅผ ๋งŒ๋“œ๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.

  • SfM initializes the model with a carefully selected two-view reconstruction.
  • Choosing a suitable initial pair is critical, since the reconstruction may never recover from a bad initialization.
  • reconstruction performance์˜ robustness, accuracy, performance๋Š” incremental process์˜ seed location์— ์˜ํ•ด ์ขŒ์šฐ๋ฉ๋‹ˆ๋‹ค.
  • image graph์—์„œ ๋งŽ์€ overlapping cameras๋ฅผ ๊ฐ€์ง€๋Š” dense location์—์„œ initailize๋ฅผ ํ•˜๋ฉด, increased redundancy์— ์˜ํ•ด ๋” robustํ•˜๊ณ  accurateํ•œ reconstruction์ด ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค.
  • ์ฆ‰, ์นด๋ฉ”๋ผ ์ค‘์ฒฉ์ด ๋งŽ์€ ๊ณณ์—์„œ image graph ์˜ denseํ•œ ์œ„์น˜์—์„œ ์ดˆ๊ธฐํ™”ํ•˜๋ฉด ์ค‘๋ณต์„ฑ ์ฆ๊ฐ€๋กœ ์„ฑ๋Šฅ์ด ์ข‹์•„์ง‘๋‹ˆ๋‹ค.
  • ๋ฐ˜๋ฉด์— ์นด๋ฉ”๋ผ๊ฐ€ ๋งŽ์ด ๊ฒน์น˜์ง€ ์•Š๋Š” sparser location์—์„œ initialize๋ฅผ ํ•˜๋ฉด, BAs์—์„œ ๋‹ค๋ค„์•ผํ•  points๊ฐ€ ์ ์–ด์ ธ์„œ sparser problems์ด ๋˜๋ฏ€๋กœ reconstruction process์˜ runtime์ด ์งง์•„์ง‘๋‹ˆ๋‹ค.

2. Image Registration: ๊ธฐํ•˜ํ•™์  ์žฌ๊ตฌ์„ฑ์œผ๋กœ ์‹œ์ž‘ํ•˜์—ฌ, ์ƒˆ๋กœ์šด ์ด๋ฏธ์ง€๋Š” ์ด๋ฏธ ๋“ฑ๋ก๋œ ์ด๋ฏธ์ง€์˜ ์‚ผ๊ฐ ์ธก๋Ÿ‰๋œ ์ ์— ๋Œ€ํ•œ ํŠน์ง• ๋Œ€์‘์„ ์‚ฌ์šฉํ•˜์—ฌ Perspective-n-Point (PnP) ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•จ์œผ๋กœ์จ ํ˜„์žฌ ๋ชจ๋ธ์— ๋“ฑ๋ก๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค (2D-3D ๋Œ€์‘).

  • PnP ๋ฌธ์ œ๋Š” ํฌ์ฆˆ $\mathbf{P}_c $์™€, uncalibrated ์นด๋ฉ”๋ผ์˜ ๊ฒฝ์šฐ, ๊ทธ ๋‚ด๋ถ€ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ์ถ”์ •ํ•˜๋Š” ๊ฒƒ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค.
  • ๋”ฐ๋ผ์„œ ์ง‘ํ•ฉ $\mathcal{P} $๋Š” ์ƒˆ๋กœ ๋“ฑ๋ก๋œ ์ด๋ฏธ์ง€์˜ ํฌ์ฆˆ $\mathbf{P}_c $์— ์˜ํ•ด ํ™•์žฅ๋ฉ๋‹ˆ๋‹ค.
  • 2D-3D ๋Œ€์‘์ด ์ข…์ข… outlier๋กœ ์˜ค์—ผ๋˜์–ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์—, calibrated ์นด๋ฉ”๋ผ์˜ ํฌ์ฆˆ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ RANSAC๊ณผ a minimal pose solver๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ถ”์ •๋ฉ๋‹ˆ๋‹ค.
  • uncalibrated ์นด๋ฉ”๋ผ์˜ ๊ฒฝ์šฐ, ๋‹ค์–‘ํ•œ minimal solvers ๋˜๋Š” sampling-based approaches๊ฐ€ ์กด์žฌํ•ฉ๋‹ˆ๋‹ค.
  • ์šฐ๋ฆฌ๋Š” ์ •ํ™•ํ•œ ํฌ์ฆˆ ์ถ”์ •๊ณผ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ์‚ผ๊ฐ ์ธก๋Ÿ‰์„ ์œ„ํ•ด ์ƒˆ๋กœ์šด ๊ฒฌ๊ณ ํ•œ ๋‹ค์Œ ์ตœ์  ์ด๋ฏธ์ง€ ์„ ํƒ ๋ฐฉ๋ฒ•์„ Sec. 4.2์—์„œ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค.

3. Triangulation: ์ƒˆ๋กœ์šด ์ด๋ฏธ์ง€๋“ค์ด ๋“ฑ๋ก๋  ๋•Œ, 3D ์”ฌ ํฌ์ธํŠธ๋“ค์ด ์‚ผ๊ฐ์ธก๋Ÿ‰์„ ํ†ตํ•ด ๊ณ„์‚ฐ๋ฉ๋‹ˆ๋‹ค. ์‚ผ๊ฐ์ธก๋Ÿ‰์€ ๋‘ ๊ฐœ ์ด์ƒ์˜ ์ด๋ฏธ์ง€๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ 3D ๊ณต๊ฐ„์˜ ์ ์„ ๊ณ„์‚ฐํ•˜๋Š” ๊ธฐ๋ฒ•์ž…๋‹ˆ๋‹ค.

  • ์ƒˆ๋กœ ๋“ฑ๋ก๋œ ์ด๋ฏธ์ง€๋Š” ๊ธฐ์กด์˜ ์žฅ๋ฉด ์ ๋“ค์„ ๊ด€์ฐฐํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
  • ๋˜ํ•œ, ์‚ผ๊ฐ ์ธก๋Ÿ‰์„ ํ†ตํ•ด ์  ์ง‘ํ•ฉ $\mathcal{X}$์„ ํ™•์žฅํ•จ์œผ๋กœ์จ scene coverage๋ฅผ ์ฆ๊ฐ€์‹œํ‚ฌ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค.
  • ์ƒˆ๋กœ์šด ์žฅ๋ฉด ์  $\mathbf{X}_k$๋Š” ๋‹ค๋ฅธ ์‹œ์ ์—์„œ ์ƒˆ๋กœ์šด ์žฅ๋ฉด ๋ถ€๋ถ„์„ ์ปค๋ฒ„ํ•˜๋Š” ์ถ”๊ฐ€ ์ด๋ฏธ์ง€๊ฐ€ ์ตœ์†Œ ํ•˜๋‚˜ ์ด์ƒ ๋“ฑ๋ก๋˜๋ฉด ์‚ผ๊ฐ ์ธก๋Ÿ‰๋˜์–ด $\mathcal{X}$์— ์ถ”๊ฐ€๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
  • ์‚ผ๊ฐ ์ธก๋Ÿ‰์€ ์ค‘๋ณต์„ฑ์„ ํ†ตํ•ด ๊ธฐ์กด ๋ชจ๋ธ์˜ ์•ˆ์ •์„ฑ์„ ์ฆ๊ฐ€์‹œํ‚ค๊ณ  ์ถ”๊ฐ€์ ์ธ 2D-3D ๋Œ€์‘์„ ์ œ๊ณตํ•จ์œผ๋กœ์จ ์ƒˆ๋กœ์šด ์ด๋ฏธ์ง€ ๋“ฑ๋ก์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๊ธฐ ๋•Œ๋ฌธ์— SfM์—์„œ ์ค‘์š”ํ•œ ๋‹จ๊ณ„์ž…๋‹ˆ๋‹ค.
  • ๋‹ค์ค‘ ๋ทฐ ์‚ผ๊ฐ ์ธก๋Ÿ‰์„ ์œ„ํ•œ ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•๋“ค์ด ์กด์žฌํ•ฉ๋‹ˆ๋‹ค.
  • ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•๋“ค์€ SfM์—์„œ ์‚ฌ์šฉํ•˜๊ธฐ์—๋Š” ์ œํ•œ๋œ robustness ๋˜๋Š” high computational cost์˜ ๋ฌธ์ œ๋ฅผ ๊ฒช๊ณ  ์žˆ์œผ๋ฉฐ, ์šฐ๋ฆฌ๋Š” Sec. 4.3์—์„œ ๊ฐ•์ธํ•˜๊ณ  ํšจ์œจ์ ์ธ ์‚ผ๊ฐ ์ธก๋Ÿ‰ ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค.

Image registration๊ณผ triangulation์€ ๋ณ„๊ฐœ์˜ ์ ˆ์ฐจ์ด์ง€๋งŒ, ๊ทธ ๊ฒฐ๊ณผ๋ฌผ์€ ๋งค์šฐ ๋ฐ€์ ‘ํ•˜๊ฒŒ ์—ฐ๊ด€๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.

  • ์นด๋ฉ”๋ผ ํฌ์ฆˆ์˜ ๋ถˆํ™•์‹ค์„ฑ์€ triangulated ์ ๋“ค์— ์ „ํŒŒ๋˜๊ณ ,
  • ๋ฐ˜๋Œ€๋กœ triangulated ์ ๋“ค์˜ ๋ถˆํ™•์‹ค์„ฑ๋„ ์นด๋ฉ”๋ผ ํฌ์ฆˆ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋ฉฐ,
  • ์ถ”๊ฐ€์ ์ธ triangulations์€ ์ค‘๋ณต์„ฑ์„ ์ฆ๊ฐ€์‹œ์ผœ ์ดˆ๊ธฐ ์นด๋ฉ”๋ผ ํฌ์ฆˆ๋ฅผ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
  • further refinement๊ฐ€ ์—†์œผ๋ฉด, SfM์€ ๋ณดํ†ต ๋น ๋ฅด๊ฒŒ ๋ณต๊ตฌ ๋ถˆ๊ฐ€๋Šฅํ•œ ์ƒํƒœ๋กœ driftํ•ฉ๋‹ˆ๋‹ค.
  • BA๋Š” futher refinement๋ฅผ ํ•ด์ค๋‹ˆ๋‹ค.

4. Bundle Adjustment, BA: ์ „์ฒด ์žฌ๊ตฌ์„ฑ ๊ฒฐ๊ณผ๋ฅผ ์ตœ์ ํ™”ํ•˜์—ฌ ์ •ํ™•๋„๋ฅผ ๋†’์ด๋Š” ๊ณผ์ •์ž…๋‹ˆ๋‹ค. BA๋Š” ์—ฌ๋Ÿฌ ๋ทฐ๋กœ๋ถ€ํ„ฐ ์–ป์–ด์ง„ ํฌ์ธํŠธ๋“ค์˜ ์œ„์น˜์™€ ์นด๋ฉ”๋ผ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ๋™์‹œ์— ์กฐ์ •ํ•˜์—ฌ ์˜ค๋ฅ˜๋ฅผ ์ตœ์†Œํ™”ํ•ฉ๋‹ˆ๋‹ค.

  • BA๋Š” ์นด๋ฉ”๋ผ ํŒŒ๋ผ๋ฏธํ„ฐ(ํฌ์ฆˆ) $\mathbf{P}_c$์™€ ์  ํŒŒ๋ผ๋ฏธํ„ฐ $\mathbf{X}_k$๋ฅผ ๋™์‹œ์— non-linear refinement๋ฅผ ํ•˜์—ฌ reprojection error(์žฌํˆฌ์˜ ์˜ค์ฐจ)๋ฅผ ์ตœ์†Œํ™”ํ•˜๋Š” ๊ธฐ๋ฒ•์ž…๋‹ˆ๋‹ค.
\[\mathit{E} = \Sigma_j \rho_j (||\pi(\mathrm{P}_c, \mathrm{X}_k - x_j||^2_2)\]
  • $\pi$: scene points๋ฅผ image space๋กœ projectํ•˜๋Š” function.
  • $\rho_k$: outliers๋ฅผ ์ž ์žฌ์ ์œผ๋กœ down-weight ํ•˜๊ธฐ ์œ„ํ•œ loss function.

image

BA๋Š” ์ด์ „ ๋‹จ๊ณ„์—์„œ ๊ณ„์‚ฐํ•œ camera pose ์™€ 3D points๋ฅผ refine ํ•˜์—ฌ reprojection error ๋ฅผ ์ตœ์†Œํ™”ํ•˜๋Š” ์ž‘์—…์ž…๋‹ˆ๋‹ค.

3D point(landmark)์—์„œ ์นด๋ฉ”๋ผ ์ขŒํ‘œ๊ณ„์˜ ์›์ ๊นŒ์ง€์˜ ์ง์„ ์„ ray๋กœ ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ๊ณ , ์ด๋Ÿฌํ•œ ray๊ฐ€ ์—ฌ๋Ÿฌ๊ฐœ ์žˆ์„ ๋•Œ โ€˜bundles of rayโ€™ ๋ผ๊ณ  ํ‘œํ˜„ํ•ฉ๋‹ˆ๋‹ค.

Reprojectionํ•œ ์œ„์น˜์™€ ์ด๋ฏธ์ง€ ์ƒ์˜ ์œ„์น˜์˜ ์ฐจ์ด์ธ reprojection error ๋ฅผ cost function์œผ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์ ํ™”๋ฅผ ์ง„ํ–‰ํ•˜๋Š”๋ฐ, image projection ๊ณผ์ •์€ non-linear ํ•˜๊ธฐ ๋•Œ๋ฌธ์— Gauss-Newton ๋“ฑ์˜ non-linear optimization ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.


Challenge

ํ˜„์žฌ ์ตœ์ฒจ๋‹จ SfM ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ๋Œ€๊ทœ๋ชจ Internet photo collections์—์„œ ๋‹ค์–‘ํ•œ ๋ณต์žกํ•œ ์ด๋ฏธ์ง€ ๋ถ„ํฌ๋ฅผ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, completeness์™€ robustness ์ธก๋ฉด์—์„œ ์™„์ „ํžˆ ๋งŒ์กฑ์Šค๋Ÿฌ์šด ๊ฒฐ๊ณผ๋ฅผ ์ž์ฃผ ๋‚ด์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค. ์ข…์ข… ์‹œ์Šคํ…œ์€ ๊ฒฝํ—˜์ ์œผ๋กœ ๋“ฑ๋ก ๊ฐ€๋Šฅํ•ด์•ผ ํ•˜๋Š” ๋งŽ์€ ์ด๋ฏธ์ง€๋ฅผ ๋“ฑ๋กํ•˜์ง€ ๋ชปํ•˜๊ฑฐ๋‚˜, mis-registrations ๋˜๋Š” drift๋กœ ์ธํ•ด broken models์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.

์ฒซ ๋ฒˆ์งธ ์›์ธ์€ correspondence search๊ฐ€ ๋ถˆ์™„์ „ํ•œ scene graph๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ๋•Œ๋ฌธ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

์˜ˆ๋ฅผ ๋“ค์–ด, matching approximations๋กœ ์ธํ•ด ์™„์ „ํ•œ ๋ชจ๋ธ์— ํ•„์š”ํ•œ ์—ฐ๊ฒฐ์„ฑ์„ ์ œ๊ณตํ•˜์ง€ ๋ชปํ•˜๊ฑฐ๋‚˜ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ์ถ”์ •์„ ์œ„ํ•œ ์ถฉ๋ถ„ํ•œ redundancy๋ฅผ ์ œ๊ณตํ•˜์ง€ ๋ชปํ•ฉ๋‹ˆ๋‹ค.

๋‘ ๋ฒˆ์งธ ์›์ธ์€ reconstruction ๋‹จ๊ณ„์—์„œ scene structure๊ฐ€ ๋ˆ„๋ฝ๋˜๊ฑฐ๋‚˜ ๋ถ€์ •ํ™•ํ•ด์„œ images๋ฅผ registerํ•˜์ง€ ๋ชปํ•˜๋Š” ๊ฒฝ์šฐ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

Image registration๊ณผ triangulation์€ ์ƒํ˜ธ๋ณด์™„์  ๊ด€๊ณ„์— ์žˆ์–ด,

  • images๋Š” ๊ธฐ์กด scene structure์—๋งŒ register๋  ์ˆ˜ ์žˆ๊ณ ,
  • scene structure๋Š” registered images์—์„œ๋งŒ triangulate๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

Incremental reconstruction ๊ณผ์ •์—์„œ ๊ฐ ๋‹จ๊ณ„์—์„œ accuracy์™€ completeness๋ฅผ ๊ทน๋Œ€ํ™”ํ•˜๋Š” ๊ฒƒ์ด SfM์˜ ์ฃผ์š” ๊ณผ์ œ์ž…๋‹ˆ๋‹ค.

๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ์ด ๊ณผ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ณ  ํ˜„์žฌ ์ตœ์ฒจ๋‹จ ๊ธฐ์ˆ ์— ๋น„ํ•ด completeness, robustness, accuracy ์ธก๋ฉด์—์„œ ๊ฒฐ๊ณผ๋ฅผ ํฌ๊ฒŒ ๊ฐœ์„ ํ•˜๋ฉด์„œ ํšจ์œจ์„ฑ์„ ๋†’์˜€์Šต๋‹ˆ๋‹ค (Sec. 5).


Contribution

๋ณธ ํŽ˜์ดํผ์—์„œ ์ฃผ์žฅํ•˜๊ณ ์ž ํ•˜๋Š” ๋ฐ”๋Š”, SfM์˜ main challenges๋ฅผ ํ‘ธ๋Š” ์ƒˆ๋กœ์šด ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์ œ์•ˆํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.

ํฌ๊ฒŒ 2๊ฐ€์ง€๋กœ ๋‚˜๋ˆ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

  • Corresspondence search๋ฅผ ์ž˜ํ•˜๊ฒŒ ํ•˜์—ฌ Scene Graph๋ฅผ ๋ณด๊ฐ•ํ•˜์—ฌ, Corresspondence search์˜ ๋‹ค์Œ ์Šคํ…์ธ Incremental Reconstruction์— ์ข‹์€ Scene Graph๋ฅผ ๋„˜๊ฒจ์ค„ ์ˆ˜ ์žˆ๋„๋ก ํ•˜๊ฑฐ๋‚˜
  • Scene Graph์— ์ถ”๊ฐ€๋˜๋Š” ์ƒˆ๋กœ์šด ์ด๋ฏธ์ง€์˜ Image Registration๊ณผ Traingulation์„ ์ž˜ํ•˜๊ฒŒ ํ•˜๊ฑฐ๋‚˜

๋ณธ ํŽ˜์ดํผ์—์„œ๋Š” ์ด์— ๋Œ€ํ•œ ๋‹ค์Œ 5๊ฐ€์ง€ Contribution์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค.

Scene Graph Augmentation, Next Best View Selection, Robust and Efficient Triangulation, Bundle Adjustment, Redundant View Mining

  1. Scene Graph Augmentation: ์ดˆ๊ธฐํ™” ๋ฐ triangulation ๊ตฌ์„ฑ ์š”์†Œ์˜ robustness๋ฅผ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ์ •๋ณด๋ฅผ ์ถ”๊ฐ€ํ•˜์—ฌ scene graph๋ฅผ ๋ณด๊ฐ•ํ•˜๋Š” geometric verification ์ „๋žต์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค.

  2. Next Best View Selection: incremental reconstruction ๊ณผ์ •์˜ robustness์™€ accuracy๋ฅผ ์ตœ๋Œ€ํ™”ํ•˜๋Š” next best view selection ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค.

  3. Robust and Efficient Triangulation: ๊ธฐ์กด ์ตœ์ฒจ๋‹จ ๊ธฐ๋ฒ•๋ณด๋‹ค ๋‚ฎ์€ ๊ณ„์‚ฐ ๋น„์šฉ์œผ๋กœ ํ›จ์”ฌ ๋” completeํ•œ scene structure๋ฅผ ์ƒ์„ฑํ•˜๋Š” robust triangulation ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค.

  4. Bundle Adjustment: drift ํšจ๊ณผ๋ฅผ ์™„ํ™”ํ•˜์—ฌ completeness์™€ accuracy๋ฅผ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ค๋Š” iterative BA, retriangulation, ๋ฐ outlier filtering ์ „๋žต์ž…๋‹ˆ๋‹ค.

  5. Redundant View Mining: redundant view mining์„ ํ†ตํ•œ dense photo collections์— ๋Œ€ํ•œ ๋” ํšจ์œจ์ ์ธ BA parameterization์ž…๋‹ˆ๋‹ค.

    • redundant view mining์€ 3D ์žฌ๊ตฌ์„ฑ ์ž‘์—…์—์„œ ์ค‘๋ณต๋œ ๋ทฐ(์ด๋ฏธ์ง€)๋ฅผ ์ฐพ์•„๋‚ด๊ณ  ํ™œ์šฉํ•˜๋Š” ๊ณผ์ •์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.
    • ์ด๋Š” robustness์™€ completeness ์ธก๋ฉด์—์„œ ํ˜„์žฌ ์ตœ์ฒจ๋‹จ ๊ธฐ๋ฒ•์„ ๋ช…ํ™•ํžˆ ๋Šฅ๊ฐ€ํ•˜๋ฉด์„œ๋„ ํšจ์œจ์„ฑ์„ ์œ ์ง€ํ•˜๋Š” ์‹œ์Šคํ…œ์„ ๊ฒฐ๊ณผ๋กœ ๋„์ถœํ•ฉ๋‹ˆ๋‹ค.

Reference

Leave a comment