1 minute read

3dgs ์•Œ๊ณ ๋ฆฌ์ฆ˜๊ณผ ์žฅ๋ฉด ํฌ๊ธฐ ๋ฐ ๋ณต์žก์„ฑ

3dgs ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ์žฅ๋ฉด์˜ ํฌ๊ธฐ(scene size)์™€๋Š” ๋…๋ฆฝ์ ์ด์ง€๋งŒ, ์žฅ๋ฉด์˜ ๋ณต์žก์„ฑ(scene complexity)๊ณผ๋Š” ๋…๋ฆฝ์ ์ด์ง€ ์•Š์Šต๋‹ˆ๋‹ค.

๋™์ผํ•œ ์žฅ๋ฉด(scene)๊ณผ ์นด๋ฉ”๋ผ(cameras)๋ฅผ ๊ณต๊ฐ„์ ์œผ๋กœ ํ™•์žฅํ•ด๋„ ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ๋™์ผํ•˜๊ฒŒ ์ž‘๋™ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.

๊ทธ๋Ÿฌ๋‚˜ ๋ช‡ ๊ฐœ์˜ ๊ฐ์ฒด(objects) ๋Œ€์‹  ์ „์ฒด ๋„์‹œ ๊ตฌ์—ญ(city district)์„ ์ฒ˜๋ฆฌํ•˜๋ ค๊ณ  ํ•˜๋ฉด, ๊ณต๊ฐ„ ํ•™์Šต๋ฅ (spatial learning rate)์ด ๋„ˆ๋ฌด ๋†’์„ ๊ฐ€๋Šฅ์„ฑ์ด ์žˆ์œผ๋ฉฐ, spatial learning rate๋ฅผ ์กฐ์ •ํ•  ํ•„์š”๊ฐ€ ์žˆ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค.

Spatial Learning Rate(๊ณต๊ฐ„ ํ•™์Šต๋ฅ )์ด๋ž€?

Spatial learning rate(๊ณต๊ฐ„ ํ•™์Šต๋ฅ )์€ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ๊ณต๊ฐ„์ ์ธ ์ •๋ณด, ์ฆ‰ ์žฅ๋ฉด์˜ ๋ฌผ๋ฆฌ์  ๋ฐฐ์น˜๋‚˜ ๊ตฌ์กฐ๋ฅผ ํ•™์Šตํ•˜๋Š” ์†๋„๋ฅผ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.

์ด๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ๊ณต๊ฐ„์  ํŠน์ง•์„ ์–ผ๋งˆ๋‚˜ ๋น ๋ฅด๊ฒŒ ์ดํ•ดํ•˜๊ณ  ์ตœ์ ํ™”ํ•˜๋Š”์ง€๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ์ค‘์š”ํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ์ž…๋‹ˆ๋‹ค.

์™œ Spatial Learning Rate(๊ณต๊ฐ„ ํ•™์Šต๋ฅ )์ด ์ค‘์š”ํ•œ๊ฐ€?

  • Small Scenes(์ž‘์€ ์žฅ๋ฉด): ์ž‘์€ ์žฅ๋ฉด์—์„œ๋Š” spatial learning rate์ด ๋†’์•„๋„ ๋ฌธ์ œ๊ฐ€ ๋˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ๋น„๊ต์  ๋‹จ์ˆœํ•œ ๊ตฌ์กฐ๋ฅผ ๋น ๋ฅด๊ฒŒ ํ•™์Šตํ•˜๊ณ  ์ตœ์ ํ™”ํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.
  • Large Scenes(ํฐ ์žฅ๋ฉด): ํฐ ์žฅ๋ฉด, ํŠนํžˆ ์ „์ฒด ๋„์‹œ ๊ตฌ์—ญ ๊ฐ™์€ ๋ณต์žกํ•œ ์žฅ๋ฉด์—์„œ๋Š” spatial learning rate์ด ๋„ˆ๋ฌด ๋†’์œผ๋ฉด ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ํ•™์Šต์„ ์ œ๋Œ€๋กœ ํ•˜์ง€ ๋ชปํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

๋”ฐ๋ผ์„œ ์žฅ๋ฉด์˜ ํฌ๊ธฐ(Scene Size)๋ณด๋‹ค๋Š” ๋ณต์žก์„ฑ(Scene Complexity)์ด ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์„ฑ๋Šฅ์— ๋” ํฐ ์˜ํ–ฅ์„ ๋ฏธ์นฉ๋‹ˆ๋‹ค.

๋ณต์žกํ•œ ์žฅ๋ฉด์„ ์ฒ˜๋ฆฌํ•  ๋•Œ๋Š” spatial learning rate์„ ์ ์ ˆํžˆ ์กฐ์ •ํ•˜์—ฌ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ์ตœ์ ์˜ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.

  • ๋ณต์žกํ•œ scene์— ๋Œ€ํ•ด์„œ๋Š” ๊ธฐ๋ณธ์ ์œผ๋กœ ๋‚ฎ์€ ์ดˆ๊ธฐ spatial learning rate๋กœ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, 0.001 ๋˜๋Š” ๊ทธ ์ดํ•˜์˜ ๊ฐ’์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
  • ํ•™์Šต๋ฅ  ๊ฐ์†Œ(Decay) ์ ์šฉ: ํ•™์Šต์ด ์ง„ํ–‰๋จ์— ๋”ฐ๋ผ spatial learning rate์„ ์ ์ฐจ ๊ฐ์†Œ์‹œํ‚ค๋Š” ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋งค ์—ํญ(epoch)๋งˆ๋‹ค learning rate์„ 0.9๋ฐฐ๋กœ ์ค„์ด๋Š” ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค.
  • Adaptive Learning Rate ์‚ฌ์šฉ: AdaGrad, RMSprop, Adam ๊ฐ™์€ adaptive learning rate ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ฌ์šฉํ•˜์—ฌ ํ•™์Šต ๊ณผ์ • ์ค‘์— ์ž๋™์œผ๋กœ spatial learning rate์„ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค.
  • ํ•™์Šต๋ฅ  ์Šค์ผ€์ค„๋ง: ํ•™์Šต๋ฅ ์„ ๋‹จ๊ณ„๋ณ„๋กœ ๋‚ฎ์ถ”๋Š” ์Šค์ผ€์ค„๋ง ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ํ•™์Šต ์˜ค๋ฅ˜๊ฐ€ ๊ฐ์†Œํ•˜์ง€ ์•Š์œผ๋ฉด spatial learning rate์„ ์ค„์ด๋Š” ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค.
  • Validation Set ํ™œ์šฉ: ๊ฒ€์ฆ ๋ฐ์ดํ„ฐ์…‹์„ ํ™œ์šฉํ•˜์—ฌ ์ตœ์ ์˜ spatial learning rate์„ ์ฐพ์•„๋ƒ…๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ํ•™์Šต๋ฅ ์„ ํ…Œ์ŠคํŠธํ•˜๊ณ , ๊ฐ€์žฅ ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฐ’์„ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค.

https://github.com/graphdeco-inria/gaussian-splatting/issues/67

image

image

Leave a comment