๋ณธ๋ฌธ ๋ฐ”๋กœ๊ฐ€๊ธฐ

representing scenes1

[Paper Review] NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis Abstract์ด ๋…ผ๋ฌธ์—์„œ๋Š” ๋ณต์žกํ•œ ์žฅ๋ฉด์˜ ์ƒˆ๋กœ์šด ์‹œ์ ์„ ์ƒ์„ฑํ•˜๋ฉฐ, ๊ธฐ์กด ์—ฐ๊ตฌ๋ฅผ ๋›ฐ์–ด๋„˜๋Š” ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•œ๋‹ค.์ด ๋ฐฉ๋ฒ•์€ ์†Œ์ˆ˜์˜ ์ž…๋ ฅ ๋ทฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์—ฐ์†์ ์ธ volume scene ํ•จ์ˆ˜๋ฅผ ์ตœ์ ํ™”ํ•œ๋‹ค.์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ์žฅ๋ฉด์„ fully connected (not convolution) deep network ๋กœ ํ‘œํ˜„ํ•œ๋‹ค. - ์ž…๋ ฅ: ์—ฐ์†์ ์ธ 5D ์ขŒํ‘œ => ๊ณต๊ฐ„ ์œ„์น˜(x, y, z) ์™€ ์‹œ์  ๋ฐฉํ–ฅ(θ, φ) ์„ ์ž…๋ ฅ- ์ถœ๋ ฅ: ํ•ด๋‹น ์œ„์น˜์˜ ์ฒด์  ๋ฐ€๋„(volume density) ์™€ ์‹œ์  ์ข…์† ๋ฐฉ์ถœ ๋ณต์‚ฌ๊ด‘(view-dependent emitted radiance) ์ถœ๋ ฅ- ํ•ฉ์„ฑ ๊ณผ์ •: ์นด๋ฉ”๋ผ์—์„œ ์žฅ๋ฉด์œผ๋กœ ๋ป—์–ด๋‚˜๊ฐ€๋Š” ๊ด‘์„ (rays) ์„ ๋”ฐ๋ผ 5D ์ขŒํ‘œ(๊ณต๊ฐ„์œ„์น˜์™€ ์‹œ์ ๋ฐฉํ–ฅ)๋ฅผ ์„ ํƒํ•œ ๋’ค, ํ•ด๋‹น ์ขŒํ‘œ์—์„œ ์‹ ๊ฒฝ๋ง์„.. 2024. 12. 10.