<aside> ๐Ÿ“Ž

https://chehun16.github.io/gs-quality/

</aside>

Abstract

This study proposes a method to enhance the quality of indoor 3D reconstruction based on 3D Gaussian Splatting (3DGS) using Polycam. The approach generates novel view camera poses, improves them with DIFIX, and incorporates geometry-aware loss terms to further refine reconstruction quality. The geometry-aware loss includes a perceptual loss applied only to novel views and normal and depth consistency losses applied to all views. These improvements enhance the accuracy of geometry reconstruction, strengthen multi-view consistency, and reduce artifacts in the reconstructed scenes. Experimental results show that the proposed method increases PSNR from 20.423 to 21.675 and SSIM from 0.856 to 0.862 compared to the original 3DGS.

๋ณธ ์—ฐ๊ตฌ๋Š” Polycam์„ ํ™œ์šฉํ•œ 3D Gaussian Splatting(3DGS) ๊ธฐ๋ฐ˜์˜ ์‹ค๋‚ด 3D reconstruction ํ’ˆ์งˆ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. Novel view camera pose ์ƒ์„ฑ ํ›„ DIFIX ๊ธฐ๋ฐ˜ ํ’ˆ์งˆ ํ–ฅ์ƒ๊ณผ geometry-aware loss term์„ ๋„์ž…ํ•˜์—ฌ ์žฌ๊ตฌ์„ฑ ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•˜์˜€์Šต๋‹ˆ๋‹ค. Geometry-aware loss์—๋Š” novel view์—๋งŒ ์ ์šฉ๋˜๋Š” perceptual loss์™€ ๋ชจ๋“  ๋ทฐ์— ์ ์šฉ๋˜๋Š” normal ๋ฐ depth consistency loss๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฐœ์„ ์„ ํ†ตํ•ด geometry ๋ณต์›์˜ ์ •ํ™•์„ฑ, multi-view consistency, ๊ทธ๋ฆฌ๊ณ  ์žฌ๊ตฌ์„ฑ ์žฅ๋ฉด์˜ artifact๊ฐ€ ๊ฐœ์„ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, ์ œ์•ˆํ•œ ๋ฐฉ๋ฒ•์€ ๊ธฐ์กด 3DGS ๋Œ€๋น„ PSNR์ด 20.423์—์„œ 21.675๋กœ, SSIM์ด 0.856์—์„œ 0.862๋กœ ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.

Introduction

Limitations of 3DGS

3D Gaussian Splatting(3DGS)์€ realistic 3D reconstruction ๊ธฐ๋ฒ•์ž…๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ scene์„ ๋ Œ๋”๋งํ•  ๋•Œ artifact๊ฐ€ ๋ฐœ์ƒํ•˜๊ฑฐ๋‚˜ ์ผ๋ถ€ ์˜์—ญ์ด ๋น„์–ด ๋ณด์ด๋Š” ๋ฌธ์ œ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.

์ด๋Ÿฌํ•œ ํ˜„์ƒ์˜ ์ฃผ์š” ์›์ธ์„ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ •๋ฆฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค:

  1. Multi view consistency ๋ถ€์กฑ
  2. ์ œํ•œ์ ์ธ loss term ๊ตฌ์„ฑ

Contribution

  1. Novel view camera pose ์ƒ์„ฑ
  2. ์ถ”๊ฐ€์ ์ธ Loss term ๋„์ž…
    1. Novel view์— ๋Œ€ํ•ด์„œ๋งŒ perceptual LPIPS loss ์ถ”๊ฐ€ํ•˜์—ฌ ํ”ฝ์…€ ์ •๋ณด ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๊ตฌ์กฐ์  ๋””ํ…Œ์ผ์„ ๋ณด์กดํ•˜์˜€์Šต๋‹ˆ๋‹ค.
    2. ์ „์ฒด view์— ๋Œ€ํ•ด normal consistency loss์™€ depth smoothness loss๋ฅผ ์ ์šฉํ•ด์„œ geometry reconstruction ํ’ˆ์งˆ์„ ๋†’์˜€์Šต๋‹ˆ๋‹ค.

Related Works

3DGS

<aside> ๐Ÿšจ 3DGS โ“

3D Gaussian Splatting for Real-Time Radiance Field Rendering (SIGGRAPH 2023)

3์ฐจ์› ๊ณต๊ฐ„์— ์žˆ๋Š” ์ ๋“ค์„ Gaussian ๋ถ„ํฌ๋กœ ํ‘œํ˜„ํ•˜์—ฌ 3D ๋ฐ์ดํ„ฐ์˜ ๋ฐ€๋„์™€ ํ˜•ํƒœ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์‹œ๊ฐํ™”ํ•˜๋Š” 3D reconstruction ๊ธฐ๋ฒ•์ž…๋‹ˆ๋‹ค.

</aside>