vkgs-sim / docs /paper /benchmarks.md
kimjunhan's picture
Add files using upload-large-folder tool
ceea884 verified
|
Raw History Blame Contribute Delete
16.2 kB

๋…ผ๋ฌธ ๋น„๊ตํ‘œ ์ดˆ์•ˆ โ€” ์ฃผํ–‰ ์žฅ๋ฉด 3DGSยท์‹œ๋ฎฌ๋ ˆ์ดํ„ฐ (2026-10-06 ์กฐ์‚ฌ)

CVPR ์ œ์ถœ์„ ๊ฐ€์ •ํ•œ ๋น„๊ตํ‘œ ์ดˆ์•ˆ. ์ˆ˜์น˜๋Š” ๊ฐ ๋…ผ๋ฌธ arXiv ์›๋ฌธ ํ‘œ(HTML)์—์„œ ์˜ฎ๊ฒผ๋‹ค. ์š”์•ฝ ๋„๊ตฌ๋ฅผ ๊ฑฐ์ณ ์ฝ์—ˆ์œผ๋ฏ€๋กœ ์ œ์ถœ ์ „ PDF ํ‘œ์™€ ๋ฐ˜๋“œ์‹œ ๋Œ€์กฐํ•  ๊ฒƒ (์กฐ์‚ฌ ์ค‘ ์—ด ์ด๋ฆ„์ด ์ž˜๋ชป ๋ถ™์€ ์‚ฌ๋ก€๊ฐ€ 1๊ฑด ์žˆ์—ˆ๋‹ค โ€” ReconDreamer++ 3 m/6 m). '(๋ณต์‚ฌ: X)' ๋Š” ๊ทธ ๋…ผ๋ฌธ์ด X ์˜ ํ‘œ๋ฅผ ์˜ฎ๊ฒจ ์˜จ ๊ฐ’, '(์žฌ์‹คํ–‰)' ์€ ๊ทธ ๋…ผ๋ฌธ ์ €์ž๊ฐ€ ๋‹ค์‹œ ๋Œ๋ฆฐ ๊ฐ’์ด๋‹ค. ๊ฐ™์€ ๋ชจ๋ธ๋„ ํ”„๋กœํ† ์ฝœ(ํ•ด์ƒ๋„ยท๋ถ„ํ• ยท์นด๋ฉ”๋ผยทํด๋ฆฝ ๊ธธ์ดยท๋งˆ์Šคํฌ)์— ๋”ฐ๋ผ ์ˆ˜ dB ์”ฉ ๋‹ค๋ฅด๋‹ค โ€” ์˜ˆ: Street Gaussians ์žฌ๊ตฌ์„ฑ PSNR ์€ OmniRe ํ‘œ์—์„œ 29.08, PVG ํ‘œ 4 ์—์„œ 32.18. ๋‹ค๋ฅธ ํ”„๋กœํ† ์ฝœ์˜ ์ˆ˜์น˜๋ฅผ ํ•œ ์—ด์— ์„ž์ง€ ์•Š๋Š”๋‹ค.


1. nuScenes โ€” ์›๋ž˜ ๊ฒฝ๋กœ (๋ณด๊ฐ„ ์‹œ์ )

1a. SplatAD ํ”„๋กœํ† ์ฝœ โ† ์šฐ๋ฆฌ๊ฐ€ ๋งž์ถœ ๊ธฐ์ค€ (์ถ”์ฒœ)

SplatAD (arXiv 2411.16816, CVPR'25) ํ‘œ 1. 10๊ฐœ ์žฅ๋ฉด 0039ยท0054ยท0061ยท0066ยท0104ยท0108ยท0122ยท0176ยท0180ยท0193, 6์บ  ์› ํ•ด์ƒ๋„(1600ร—900), ํ•œ ํ”„๋ ˆ์ž„ ๊ฑธ๋Ÿฌ ํ•™์Šต / ๋‚˜๋จธ์ง€ ํ‰๊ฐ€. ๊ธฐ์ค€์„  ์ „๋ถ€ SplatAD ์ €์ž ์žฌ์‹คํ–‰ (๊ณต๊ฐœ ์ฝ”๋“œ neurad-studio).

๋ฐฉ๋ฒ• PSNR SSIM LPIPS
UniSim 22.66 0.743 0.414
PVG 24.94 0.768 0.497
OmniRe 25.50 0.798 0.386
Street Gaussians 25.54 0.799 0.375
NeuRAD 26.17 0.790 0.312
SplatAD 27.54 0.849 0.302
  • 10๊ฐœ ์ค‘ 6๊ฐœ๊ฐ€ ์‹ฑ๊ฐ€ํฌ๋ฅด (0039ยท0054ยท0061ยท0122ยท0193 one-north, 0180 queenstown โ€” ์šฐ๋ฆฌ ์ธ๋ฑ์Šค์— ์žˆ์Œ), 4๊ฐœ(0066ยท0104ยท0108ยท0176)๋Š” ๋ณด์Šคํ„ด. โ†’ ๋ณด์Šคํ„ด 4๊ฐœ๋ฅผ ์ธ๋ฑ์Šค์— ๋”ํ•ด 10๊ฐœ ์ „์ฒด๋ฅผ ๋Œ๋ฆฌ๊ฑฐ๋‚˜, '์‹ฑ๊ฐ€ํฌ๋ฅด 6๊ฐœ ๋ถ€๋ถ„์ง‘ํ•ฉ' ์œผ๋กœ ๋”ฐ๋กœ ํ‘œ๊ธฐ (๊ธฐ์ค€์„ ๋„ ๊ฐ™์€ 6๊ฐœ๋กœ ๋‹ค์‹œ ๊ณ„์‚ฐํ•ด์•ผ ํ•จ โ€” ๋…ผ๋ฌธ์€ ์žฅ๋ฉด๋ณ„ ์ˆ˜์น˜๋ฅผ ์ฃผ์ง€ ์•Š์Œ, ํ™•์ธ ํ•„์š”).
  • 'ํ”„๋ ˆ์ž„' ์ด 2 Hz ํ‚คํ”„๋ ˆ์ž„์ธ์ง€ 12 Hz sweep ํฌํ•จ์ธ์ง€ ๋…ผ๋ฌธ์— ๋ช…์‹œ ์—†์Œ โ€” ํ™•์ธ ํ•„์š” (๋ถ„ํ•  ๋‚œ์ด๋„๊ฐ€ ๋‹ฌ๋ผ์ง„๋‹ค).

1b. ๊ทธ ๋ฐ–์˜ nuScenes ํ”„๋กœํ† ์ฝœ (์ฐธ๊ณ  โ€” ์šฐ๋ฆฌ์™€ ์ง์ ‘ ๋น„๊ต ๋ถˆ๊ฐ€)

์ถœ์ฒ˜ ์„ค์ • ๋ฐฉ๋ฒ•: PSNR / SSIM / LPIPS
DrivingGaussian (2312.07920) ํ‘œ 1 6๊ฐœ ์žฅ๋ฉด(์ด๋ฆ„ ๋น„๊ณต๊ฐœ), 6์บ  1600ร—900, 5์žฅ๋งˆ๋‹ค ํ‰๊ฐ€ 3D-GS 26.08/.717/.298 ยท EmerNeRF 26.75/.760/.311 ยท S-NeRF 25.43/.730/.302 ยท DrivingGaussian-L 28.74/.865/.237
NeuRAD (2311.15260) ํ‘œ 1 (S-NeRF ํ”„๋กœํ† ์ฝœ) 0164ยท0209ยท0359ยท0916, 6์บ , 40 ์ƒ˜ํ”Œ ์ค‘ 4์žฅ๋งˆ๋‹ค ํ‰๊ฐ€ Mip-NeRF360* 24.37/.795/.240 ยท S-NeRF 26.21/.831/.228 ยท NeuRAD 26.99/.815/.225 ยท NeuRAD-2x 27.13/.820/.205
ReconDrive (2603.07552, 2026) ํ‘œ 1 ๊ฒ€์ฆ 14์žฅ๋ฉด, 6์บ  518ร—280, 0.5 s ๋งˆ๋‹ค 1์žฅ๋งŒ ํ•™์Šต(ํฌ์†Œ) ์ƒˆ ์‹œ์ : Street GS 22.98/.696/.295 ยท PVG 23.48/.692/.290 ยท OmniRe 23.01/.689/.276 ยท ReconDrive(ํ”ผ๋“œํฌ์›Œ๋“œ) 23.99/.723/.259
Nighttime PBR-GS (2602.13549) ํ‘œ II ์•ผ๊ฐ„ 11์žฅ๋ฉด, 8์žฅ๋งˆ๋‹ค ํ‰๊ฐ€ ์ƒˆ ์‹œ์ : 3DGS 26.4 ยท Street GS 26.9 ยท OmniRe 26.9 ยท ์ œ์•ˆ 27.7
HUGSIM (2412.01718) ํ‘œ V 2๊ฐœ ์žฅ๋ฉด(๋น„๊ณต๊ฐœ), 5์žฅ๋งˆ๋‹ค ์ œ์™ธ HUGSIM 25.36/.776/.171 (KID .062) ยท NeuRAD 24.49/.692/.287 (KID .082)
ReconDreamer++ (2503.18438) ํ‘œ 2 ์›๋ž˜ ๊ฒฝ๋กœ Street GS 34.91/.9528/.1185 ยท ReconDreamer 33.85 ยท ReconDreamer++ 34.69/.9533/.1161

์ž์ฒด ๋…ผ๋ฌธ์— nuScenes ์ˆ˜์น˜๊ฐ€ ์—†๋Š” ๋ชจ๋ธ: OmniRe(์ •์„ฑ๋งŒ), SยณGaussian, EmerNeRF, Street Gaussians, PVG โ€” nuScenes ์ˆ˜์น˜๋Š” ์œ„์ฒ˜๋Ÿผ ๋‚จ์ด ์žฌ์‹คํ–‰ํ•œ ๊ฒƒ๋งŒ ์žˆ๋‹ค.

LiDAR (SplatAD ํ‘œ 2, nuScenes): SplatAD ์ค‘์•™๊ฐ’ ๊นŠ์ด ์˜ค์ฐจ 0.02 / Chamfer 1.7, NeuRAD 0.01 / 6.3.

1c. ์šฐ๋ฆฌ ํ˜„์žฌ ์ˆ˜์น˜ (๋น„๊ต ๋ถˆ๊ฐ€ ์กฐ๊ฑด โ€” ํ‘œ์— ๋„ฃ์ง€ ๋ง ๊ฒƒ)

scene-0003 ํ•˜๋‚˜, 6์บ  800ร—450, 8์žฅ๋งˆ๋‹ค ํ‰๊ฐ€, ์ •์  ์˜์—ญ๋งŒ PSNR: 25.0 dB / SSIM 0.79 (ํ™˜๊ฒฝ ์ „์šฉ ๋ชจ๋ธ). SplatAD ํ”„๋กœํ† ์ฝœ๋กœ ๋ฐ”๊พธ๋ ค๋ฉด: โ‘  10(๋˜๋Š” 6)๊ฐœ ์žฅ๋ฉด โ‘ก ํ•œ ์žฅ ๊ฑธ๋Ÿฌ ํ•™์Šต โ‘ข 1600ร—900 ํ‰๊ฐ€ โ‘ฃ ์ „์ฒด ์˜์ƒ PSNR (์ง€๊ธˆ์€ ์ฐจ๋Ÿ‰์„ ์ง€์šด ํ™˜๊ฒฝ ์ „์šฉ์ด๋ผ ์ „์ฒด ์˜์ƒ์—์„œ ํฌ๊ฒŒ ๋ถˆ๋ฆฌ โ€” ๋น„๊ต์šฉ์œผ๋กœ๋Š” ๋™์  ๊ฐ์ฒด๋ฅผ ๋ชจ๋ธ๋งํ•˜๋Š” ํ•™์Šต์ด ๋”ฐ๋กœ ํ•„์š”) โ‘ค ๊ฐ™์€ ํ”„๋ ˆ์ž„ ์†Œ์Šค.


2. Waymo โ€” ์›๋ž˜ ๊ฒฝ๋กœ

2a. OmniRe ํ”„๋กœํ† ์ฝœ (2408.16760 ํ‘œ 1) โ€” ๋™์  ์žฅ๋ฉด ๋น„๊ต์˜ ์‚ฌ์‹ค์ƒ ํ‘œ์ค€

๋™์  8 ์„ธ๊ทธ๋จผํŠธ(์‚ฌ๋žŒยท์ž์ „๊ฑฐ ๋งŽ์Œ), ์•ฝ 150 ํ”„๋ ˆ์ž„, ์•ž 3์บ  640ร—960, 10์žฅ๋งˆ๋‹ค ํ‰๊ฐ€. ๊ธฐ์ค€์„ ์€ OmniRe ์ €์ž๊ฐ€ LiDAR ๊นŠ์ดยทAbsGrad ๋ฅผ ๋„ฃ์–ด ์žฌ์‹คํ–‰. ๊ฐ’์€ PSNR/SSIM.

๋ฐฉ๋ฒ• ์žฌ๊ตฌ์„ฑ ์ „์ฒด ์žฌ๊ตฌ์„ฑ ์‚ฌ๋žŒ ์žฌ๊ตฌ์„ฑ ์ฐจ๋Ÿ‰ ์ƒˆ ์‹œ์  ์ „์ฒด ์ƒˆ ์‹œ์  ์‚ฌ๋žŒ ์ƒˆ ์‹œ์  ์ฐจ๋Ÿ‰
3DGS 26.00/.912 16.88/.414 16.18/.425 25.57/.906 16.62/.387 16.00/.407
DeformableGS 28.40/.929 17.80/.460 19.53/.570 27.72/.922 17.30/.426 18.91/.530
HUGS 28.26/.923 16.23/.404 24.31/.794 27.65/.914 15.99/.378 23.27/.748
Street GS 29.08/.936 16.83/.420 27.73/.880 28.54/.928 16.55/.393 26.71/.846
EmerNeRF 31.93/.902 22.88/.578 24.65/.723 29.67/.883 20.32/.454 22.07/.609
PVG 32.37/.937 24.06/.703 25.02/.787 30.19/.919 21.30/.567 22.28/.679
OmniRe 34.25/.954 28.15/.845 28.91/.892 32.57/.942 24.36/.727 27.57/.858

OmniRe ํ‘œ 3 (LiDAR ๊นŠ์ด, Chamfer/RMSE, ์ƒˆ ํ”„๋ ˆ์ž„): 3DGS .467/2.896 ยท StreetGS .286/2.228 ยท OmniRe .244/1.909. OmniRe ํ‘œ 5 (์ž์ฒด 32 ๋™์  ์„ธ๊ทธ๋จผํŠธ โ€” NOTR dynamic-32 ์™€ ๋‹ค๋ฅธ ์ง‘ํ•ฉ์œผ๋กœ ๋ณด์ž„, ํ™•์ธ ํ•„์š”): ์ƒˆ ์‹œ์  ์ „์ฒด EmerNeRF 29.04 ยท StreetGS 28.73 ยท OmniRe 31.71. DeSiRe-GS (2411.11921) ํ‘œ 2: ๊ฐ™์€ ํ”„๋กœํ† ์ฝœ ์žฌ๊ตฌ์„ฑ 33.82 / ์ƒˆ ์‹œ์  31.49 (PSNR ๋งŒ).

2b. NOTR (EmerNeRF ํ”„๋กœํ† ์ฝœ) โ€” ์•ž 3์บ  640ร—960, 10์žฅ๋งˆ๋‹ค ํ‰๊ฐ€

์ถœ์ฒ˜ ๋ฐฉ๋ฒ• ์žฌ๊ตฌ์„ฑ PSNR/SSIM/LPIPS (๋™์  D-PSNR) ์ƒˆ ์‹œ์  PSNR/SSIM/LPIPS (D-PSNR)
EmerNeRF (2311.02077) ํ‘œ 1a, dynamic-32 ์ „์ฒด ๋กœ๊ทธ EmerNeRF 28.87/.814/โ€“ (26.19) 27.62/.792/โ€“ (24.18)
SยณGaussian (2405.20323) ํ‘œ 1, dynamic, 100 ํ”„๋ ˆ์ž„ ํด๋ฆฝ 3DGS 28.47/.876/.136 (23.26) 25.14/.813/.165 (20.48)
ใ€ƒ EmerNeRF (์žฌ์‹คํ–‰) 28.16/.806/.228 (24.32) 25.14/.747/.313 (23.49)
ใ€ƒ SยณGaussian 31.35/.911/.106 (26.02) 27.44/.857/.137 (22.92)
4DGS ๋„ค์ดํ‹ฐ๋ธŒ (2412.20720) ํ‘œ 8, 50 ํ”„๋ ˆ์ž„ Street GS (์žฌ์‹คํ–‰) 29.89/.906/.125 28.19/.867/.133
ใ€ƒ 4DGS 35.04/.943/.057 28.67/.868/.102
IDSplat (2511.19235) ํ‘œ 1, DeSiRe ์„ค์ • SplatAD โ€“ 30.80/.900/.160 (28.97)
ใ€ƒ IDSplat โ€“ 30.83/.900/.160 (29.20)
SยณGaussian ํ‘œ 1, static-32 3DGS 29.42/.891/.118 26.82/.836/.134
ใ€ƒ EmerNeRF (์žฌ์‹คํ–‰) 30.00/.834/.201 28.89/.814/.212
ใ€ƒ SยณGaussian 30.73/.883/.116 27.05/.825/.142

2c. PVG ํ‘œ 4 (12 ์„ธ๊ทธ๋จผํŠธ, ์•ž 3์บ  960ร—640, ์ „๋ถ€ ์žฌ์‹คํ–‰) โ€” ์ •์ /๋™์  ๋ถ„๋ฆฌ ์ˆ˜์น˜๊ฐ€ ์žˆ์–ด ์œ ์šฉ

๊ฐ’: PSNR/SSIM/LPIPS, ๊ทธ ๋’ค ์ •์  S / ๋™์  D PSNR.

๋ฐฉ๋ฒ• ์žฌ๊ตฌ์„ฑ ์ƒˆ ์‹œ์ 
Street GS 32.18/.918/.090, S 32.68 / D 28.64 28.92/.877/.110, S 29.38 / D 25.54
HUGS 31.58/.904/.094, S 32.79 / D 25.78 29.34/.865/.110, S 30.42 / D 23.84
OmniRe 32.32/.923/.084, S 32.92 / D 28.36 29.41/.884/.101, S 29.91 / D 25.85
DeformableGS 32.34/.923/.086, S 32.94 / D 27.39 29.52/.889/.100, S 30.37 / D 24.66
4DGS 34.94/.935/.068, S 35.36 / D 31.80 26.56/.793/.132, S 26.87 / D 23.97
PVG 36.02/.952/.059, S 36.79 / D 32.17 30.95/.897/.105, S 31.72 / D 25.82

2d. Street Gaussians ํ”„๋กœํ† ์ฝœ (2401.01339 ํ‘œ 1) โ€” 8 ์„ธ๊ทธ๋จผํŠธ, ์•ž ์นด๋ฉ”๋ผ 1066ร—1600, 4์žฅ๋งˆ๋‹ค ํ‰๊ฐ€ (์ƒˆ ์‹œ์ )

3DGS 29.64/.918/.117 ยท EmerNeRF 30.87/.905/.133 ยท MARS 29.75/.886/.264 ยท Street GS 34.61/.938/.079 (๋ฐ•์Šค ์•ˆ PSNR* 30.23). ๊ฐ™์€ 8 ์„ธ๊ทธ๋จผํŠธ 1280ร—1920 (AD-GS ์„ค์ •, IDSplat ํ‘œ 1): Street GS 33.97 ยท 4DGS 34.64 ยท SplatAD 34.24 ยท IDSplat 34.59. โ€ป SยณGaussian ์ด '๋นŒ๋ ค ์™”๋‹ค' ๋Š” Street GS ๊ธฐ์ค€์„  ์ˆ˜์น˜๋Š” ํ˜„์žฌ arXiv ํ‘œ์™€ ๋‹ค๋ฆ„ โ€” ์“ฐ์ง€ ๋ง ๊ฒƒ.


3. ์ƒˆ ์‹œ์  โ€” ์ฐจ์„  ์ด๋™ (์šฐ๋ฆฌ ์ ์ง„ Difix ์ฆ๋ฅ˜ ์ฃผ์žฅ์˜ ํ‰๊ฐ€)

3a. ์ •๋‹ต ์—†๋Š” ์ด๋™ (๋ถ„ํฌยท์ผ๊ด€์„ฑ ์ง€ํ‘œ)

์ถœ์ฒ˜ ๋ฐ์ดํ„ฐ ๋ฐฉ๋ฒ• ๊ฒฐ๊ณผ
FreeVS (2410.18079) ํ‘œ 2 Waymo 12 ์‹œํ€€์Šค, ์˜† 0/1/2/4 m 3DGS ยท EmerNeRF ยท Street GS ยท FreeVS FID@1m 52.07 ยท 58.26 ยท 41.17 ยท 13.45; FID@2m 61.16 ยท 69.50 ยท 55.71 ยท 16.60; FID@4m 86.21 ยท 84.81 ยท 80.44 ยท 22.08
StreetCrafter (2412.13188) ํ‘œ 1 Waymo 15 ์‹œํ€€์Šค, ์˜† 2/3 m 3DGS ยท Street GS ยท StreetCrafter(3DGS ์ฆ๋ฅ˜) FID@2m 100.50 ยท 71.42 ยท 58.17; FID@3m 122.52 ยท 93.38 ยท 71.40; ๋ณด๊ฐ„ PSNR 28.85 ยท 30.95 ยท 30.05; FPS 213 ยท 92 ยท 113
StreetCrafter ํ‘œ 2 PandaSet 5 ์‹œํ€€์Šค UniSim ยท NeuRAD ยท Street GS ยท StreetCrafter(์ฆ๋ฅ˜) FID@2m 75.26 ยท 64.65 ยท 69.87 ยท 62.15; FID@3m 92.65 ยท 86.44 ยท 90.41 ยท 78.88
ReconDreamer (2411.19548) ํ‘œ 1 Waymo 8 ์„ธ๊ทธ๋จผํŠธ, ์•ž ์นด๋ฉ”๋ผ Street GS โ†’ ReconDreamer 3 m: NTA-IoU .498โ†’**.539**, NTL-IoU 53.19โ†’54.58, FID 130.75โ†’93.56; 6 m: FID 213.04โ†’149.19
ReconDreamer++ (2503.18438) nuScenes, 3 m ReconDreamer โ†’ ReconDreamer++ NTA-IoU .325โ†’**.365**, NTL-IoU 49.72โ†’51.33, FID 125.43โ†’118.27
DriveDreamer4D (2410.13571) Waymo 8 ์„ธ๊ทธ๋จผํŠธ, ์ฐจ์„  ๋ณ€๊ฒฝ PVG โ†’ +DD4D NTA-IoU .256โ†’**.438**, FID 105.29โ†’71.52
UniSim (2308.01898) PandaSet UniSim FID@2m 74.7, @3m 97.5

3b. ์ •๋‹ต ์žˆ๋Š” ์ด๋™ (์‹ค์ธก ํ‰ํ–‰ ์ฐจ์„ )

  • Para-Lane (2502.15635): ํ‰ํ–‰ 3์ฐจ์„  25 ๊ทธ๋ฃน(์ฐจ์„  ๊ฐ„๊ฒฉ 3โ€“4 m), ์•ž 1920ร—1080. ์ธ์ ‘ ์ฐจ์„  PSNR/SSIM/LPIPS: 3DGS 17.05/.524/.446 ยท Scaffold-GS 17.59 ยท Street GS 17.50/.510/.456 (๊ฐ™์€ ์ฐจ์„  ์•ฝ 23 โ†’ ํ•œ ์ฐจ์„ ์— ์•ฝ โˆ’5.5 dB).
  • EUVS (2412.05256, ICCV'25): nuPlanยทArgoverse2ยทMARS ๋‹ค์ค‘ ์ฃผํ–‰. ์„ค์ • 1(์ด๋™๋งŒ): 3DGS 16.37/.7203/.2599 ยท GSPro 16.39 (๋ฐฉ๋ฒ•๋ณ„ ํ‘œ ๋ฒˆํ˜ธ ํ™•์ธ ํ•„์š”). ReconDreamer++ ๊ฐ€ ํ•œ ์ฐจ์„  ์˜คํ”„์…‹ 20.52/.7033/.3057 ๋ณด๊ณ .
  • Difix3D+ (2503.01774) ํ‘œ 3: ์ž์ฒด RDS (๊ฐ€์šด๋ฐ ์นด๋ฉ”๋ผ๋กœ ํ•™์Šต, ์˜† ์นด๋ฉ”๋ผ ํ‰๊ฐ€ โ€” ์ฐจ์„  ์ด๋™์ด ์•„๋‹˜) Nerfacto 19.95/.4930/.5300/FID 91.38 โ†’ +Difix3D+ 21.75/.5829/.4016/73.08.
  • OneFixer (2609.32175): ์ดˆ๋ก๋งŒ ํ™•์ธ โ€” FVDยทLPIPSยทDISTS + closed-loop ์ถฉ๋Œ๋ฅ  (์›์‹œ 3DGS ๋Œ€๋น„ ์•ฝ 1/3 ๊ฐ์†Œ). ํ‘œ ์ˆ˜์น˜ ํ™•์ธ ํ•„์š”.

์šฐ๋ฆฌ ํ”„๋กœํ† ์ฝœ (์ถ”์ฒœ): ReconDreamer++ ์˜ nuScenes ํ”„๋กœํ† ์ฝœ์„ ๋”ฐ๋ฅธ๋‹ค โ€” โ‘  ์›๋ž˜ ๊ฒฝ๋กœ PSNR/SSIM/LPIPS (๋‚˜๋น ์ง€์ง€ ์•Š์Œ์„ ๋ณด์ž„) โ‘ก ์˜† 1ยท2ยท3 m ์—์„œ FIDยทKID + NTA-IoU(nuScenes 3D ๋ฐ•์Šค ํˆฌ์˜ vs ๊ฒ€์ถœ๊ธฐ) + NTL-IoU(nuScenes ๋งต ์ฐจ์„  ํˆฌ์˜ vs ์ฐจ์„  ๊ฒ€์ถœ๊ธฐ) โ‘ข ๊ธฐ์ค€์„ ์€ Street GSยทOmniReยท์›์‹œ 3DGS ๋ฅผ ์šฐ๋ฆฌ ์žฅ๋ฉด์—์„œ ์žฌ์‹คํ–‰(๋…ผ๋ฌธ ๊ฐ„ FID ๋Š” ์žฅ๋ฉดยทํ‘œ๋ณธ ์ˆ˜๊ฐ€ ๋‹ฌ๋ผ ๋น„๊ต ๋ถˆ๊ฐ€ โ€” FreeVSยทStreetCrafter ์ˆ˜์น˜๋Š” ๋งฅ๋ฝ์œผ๋กœ๋งŒ) โ‘ฃ ์ •๋‹ต ์žˆ๋Š” ์ฃผ์žฅ์ด ํ•„์š”ํ•˜๋ฉด EUVS ์„ค์ • 1 ๋˜๋Š” Para-Lane ์ธ์ ‘ ์ฐจ์„  ์ถ”๊ฐ€. ์šฐ๋ฆฌ๊ฐ€ ์ด๋ฏธ ๊ฐ€์ง„ ๊ฒƒ: ์‚ฌ์ดํด ๋ Œ๋” ๋ฒค์น˜๋งˆํฌ(OneFixer ๋ถ€๋ก A.1 ๋ฐฉ์‹, LPIPS/DISTS) โ€” scene-0003 1 m ์ด๋™ LPIPS 0.391. (+Difix 0.162 ๋Š” ๋ฌดํšจ โ€” ์ฐธ์กฐ๋กœ ์ •๋‹ต๊ณผ 0.13 m ๋–จ์–ด์ง„ ์‹ค์ œ ์‚ฌ์ง„์„ ์จ์„œ ์ •๋‹ต์„ ๋ฒ ๊ผˆ๋‹ค. 2026-10-06 ์ˆ˜์ •: ์ฐธ์กฐ๋Š” ์ด๋™ ๊ฑฐ๋ฆฌ ์ด์ƒ ๋–จ์–ด์ง„ ํ”„๋ ˆ์ž„์—์„œ. ๋‹ค์‹œ ์žด ๊ฒƒ.)


4. closed-loop ์‹œ๋ฎฌ๋ ˆ์ดํ„ฐ ์ง€ํ‘œ

ํ”„๋กœํ† ์ฝœ ์ง€ํ‘œ ์ •์˜ ๋Œ€ํ‘œ ์ˆ˜์น˜ ๋ Œ๋”๋ง
HUGSIM (2412.01718) HD-Score_t = NC ร— DAC ร— ๊ฐ€์ค‘ํ‰๊ท (TTC, COM); HD-Score = Rc ร— ์‹œ๊ฐ„ ํ‰๊ท . KITTI-360ยทWaymoยทnuScenesยทPandaSet 70+ ์‹œํ€€์Šค, 400+ ์‹œ๋‚˜๋ฆฌ์˜ค, 4 ๋‚œ์ด๋„ ์ฃผํ–‰ ์ •์ฑ… ์ ์ˆ˜๋Š” ๊ทธ๋ฆผ(Fig. 18)๋งŒ, ํ‘œ ์—†์Œ. ์ƒˆ ์‹œ์  nuScenes 25.36/.776/.171/KID .062 89.15 FPS (RTX 3090), 445๋งŒ ๊ฐ€์šฐ์‹œ์•ˆ, ํ•ด์ƒ๋„ ๋ฏธ๊ธฐ์žฌ
RAD (2502.13144) ํ‘œ 4 ์ถฉ๋Œ๋ฅ  CR, ์ดํƒˆ๋ฅ  DR, ADD, ์ €ํฌ VADv2 โ†’ RAD: CR .270โ†’.089, DR .243โ†’.063 ๋ฏธ๊ธฐ์žฌ
NeuroNCAP (2404.07762) ํ‘œ 1 NNS (์ถฉ๋Œ ์—†์œผ๋ฉด 5, ์žˆ์œผ๋ฉด ์ถฉ๋Œ ์†๋„๋กœ ๊ฐ์ ), nuScenes 14 ์‹œํ€€์Šค ร— 100ํšŒ UniAD 0.73 (์ถฉ๋Œ 88.6%) ยท VAD 0.66 (92.5%) NeuRAD ๊ธฐ๋ฐ˜
NAVSIM v1 (2406.15349) PDMS = NC ร— DAC ร— (5ยทEP + 5ยทTTC + 2ยทC)/12 TransFuser 84.0 ยท UniAD 83.4 ยท ์‚ฌ๋žŒ 94.8 ๋ Œ๋”๋ง ์—†์Œ
NAVSIM v2 (2506.04218) EPDMS (+DDC, TLC, LK, HC, EC). 2๋‹จ๊ณ„๋Š” 3DGS ๋กœ ํ•ฉ์„ฑํ•œ ๊ด€์ธก navhard: LTF 23.1 ยท PDM-Closed 51.3. 3DGS ์˜์‚ฌ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ โ†” ์‹ค์ œ closed-loop ์ƒ๊ด€ r = 0.89 ์˜คํ”„๋ผ์ธ ์‚ฌ์ „ ์ƒ์„ฑ
Bench2Drive (2406.03877) DS, SR (CARLA) UniAD-Base 45.81/16.36 ยท VAD 42.35/15.00 ๊ฒŒ์ž„ ์—”์ง„ โ€” ๋งฅ๋ฝ์œผ๋กœ๋งŒ

๋ Œ๋”๋ง ์†๋„ ๋น„๊ต (์ฃผํ–‰ 3DGS): Street GS 135 FPS (1066ร—1600, RTX 4090) ยท 3DGS 205 ยท EmerNeRF 0.21 ยท StreetCrafter ์ฆ๋ฅ˜ 113 ยท HUGSIM 89 (RTX 3090). ์šฐ๋ฆฌ: RTX 3060 Ti(8 GB)์—์„œ 1600ร—900 ํ•œ ์žฅ ์ •๋ ฌ 0.9โ€“2.7 ms + ๋ž˜์Šคํ„ฐ ์•ฝ 15 ms (GPU ๊ณต์œ  ์ƒํƒœ ์ธก์ •) โ€” 6์บ  1600ร—900 ์ „์ฒด ๋ฃจํ”„ Hz ๋ฅผ ๋”ฐ๋กœ ์žฌ์„œ GPU ๋“ฑ๊ธ‰ ์ฐจ์ด๋ฅผ ๋ฐํ˜€ ํ•จ๊ป˜ ํ‘œ๊ธฐํ•  ๊ฒƒ.

์šฐ๋ฆฌ ํ”„๋กœํ† ์ฝœ (์ถ”์ฒœ): HUGSIM ์‹ HD-Score ํ•ญ๋ชฉ(NCยทDACยทTTCยทCOMยทRc) ๋˜๋Š” NAVSIM EPDMS ํ•ญ๋ชฉ์œผ๋กœ ์ฃผํ–‰ ์ ์ˆ˜, ๋ Œ๋”๋Ÿฌ FPS ์™€ ๋ฃจํ”„ Hz ๋ฅผ ๋”ฐ๋กœ, NAVSIM v2 ์˜ r = 0.89 ๋ฅผ '3DGS ์žฌ์‹œ๋ฎฌ๋ ˆ์ด์…˜์ด ์œ ํšจํ•˜๋‹ค' ๊ทผ๊ฑฐ๋กœ ์ธ์šฉ.


5. ๋‚ ์”จ ํ•ฉ์„ฑ โ€” ํ‘œ์ค€ ๋ฒค์น˜๋งˆํฌ ์—†์Œ

์ถœ์ฒ˜ ํ‰๊ฐ€ ์ˆ˜์น˜
ClimateNeRF (2211.13226, ICCV'23) ์‚ฌ์šฉ์ž ์—ฐ๊ตฌ๋งŒ (37๋ช…, 2,664 ์Œ) ์„ ํ˜ธ ์•ฝ 80โ€“85%, ์‹ค์‹œ๊ฐ„ ์•„๋‹˜
WeatherEdit (2505.20471) CLIP-S / CLIP-DS, warp error, FID(์ ˆ์ œ), ๋‹ค์šด์ŠคํŠธ๋ฆผ ๋ถ„ํ•  mIoU CLIP-S .750 / CLIP-DS .302 (InstructPix2Pix .691/.295), ACDCยทMUSE ์ตœ๋Œ€ +14.9% mIoU, 3D ๋ Œ๋” ํ”„๋ ˆ์ž„๋‹น 0.21 s
Weather-Magician (2505.19919) CMMD, UVQ RTX 4090 1200ร—800: ๋ง‘์Œ 83 FPS ยท ์•ˆ๊ฐœ 31 ยท ๋น„/๋ˆˆ 10.4 ยท ๋ˆˆ ์Œ“์ž„ 58
GSRAIN (2608.02177) FID, ๊ฐ•์šฐ๋Ÿ‰ ์ œ์–ด 0โ€“13 mm/h FID 149.09 (CycleGAN-Turbo 155.71, WeatherEdit 157.94) โ€” ์ดˆ๋ก๋งŒ ํ™•์ธ

์šฐ๋ฆฌ ํ”„๋กœํ† ์ฝœ (์ถ”์ฒœ): WeatherEdit ์‹ CLIP-S/CLIP-DS + ์‹ค์ œ ์šฐ์ฒœ ์˜์ƒ(ACDC ๋˜๋Š” nuScenes ๋ฐคยท๋น„ ์”ฌ) ๋Œ€๋น„ FID/KID, warp error(์‹œ๊ฐ„ ์ผ๊ด€์„ฑ), ์‚ฌ์šฉ์ž ์—ฐ๊ตฌ, ๋‹ค์šด์ŠคํŠธ๋ฆผ ๊ฒ€์ถœ ์„ฑ๋Šฅ ์ €ํ•˜, ํšจ๊ณผ๋ณ„ FPS (Weather-Magician ๋Œ€๋น„ โ€” ์šฐ๋ฆฌ ๊ฐ•์ˆ˜ ์ž…์ž 0.05 ms + ๋‚ ์”จ ํŒจ์Šค 0.2โ€“0.5 ms). 6์บ  ์ผ๊ด€์„ฑ(๊ฐ™์€ ๋น—๋ฐฉ์šธ)์€ ๊ธฐ์กด ์—ฐ๊ตฌ๊ฐ€ ๋‹ค๋ฃจ์ง€ ์•Š๋Š” ์ฐจ๋ณ„์  โ€” ์นด๋ฉ”๋ผ ๊ฒน์นจ ์˜์—ญ ์ผ๊ด€์„ฑ ์ง€ํ‘œ๋ฅผ ์ƒˆ๋กœ ์ •์˜ํ•  ์ˆ˜ ์žˆ๋‹ค.


6. ์šฐ๋ฆฌ๊ฐ€ ํ•  ์ผ (๋น„๊ตํ‘œ๋ฅผ ์ฑ„์šฐ๋ ค๋ฉด)

  1. SplatAD nuScenes ํ”„๋กœํ† ์ฝœ ์žฌํ˜„ โ€” 6๊ฐœ ์‹ฑ๊ฐ€ํฌ๋ฅด ์žฅ๋ฉด(+ ๋ณด์Šคํ„ด 4๊ฐœ ์ธ๋ฑ์Šค ์ถ”๊ฐ€ ์—ฌ๋ถ€ ๊ฒฐ์ •), ํ•œ ์žฅ ๊ฑธ๋Ÿฌ ํ•™์Šต, 1600ร—900, ์ „์ฒด ์˜์ƒ PSNR. ๋™์  ๊ฐ์ฒด๋ฅผ ์ง€์šด ํ™˜๊ฒฝ ์ „์šฉ ๋ชจ๋ธ๊ณผ ๋ณ„๋„๋กœ, ๋™์  ๊ฐ์ฒด๋ฅผ ํฌํ•จํ•œ ํ•™์Šต์ด ํ•„์š”ํ•˜๋‹ค.
  2. ์ฐจ์„  ์ด๋™ ํ‰๊ฐ€ ๋„๊ตฌ โ€” FID/KID (1ยท2ยท3 m), NTA-IoU (nuScenes ๋ฐ•์Šค + ๊ฒ€์ถœ๊ธฐ), NTL-IoU (nuScenes ๋งต ์ฐจ์„  + ์ฐจ์„  ๊ฒ€์ถœ๊ธฐ).
  3. ๊ธฐ์ค€์„  ์žฌ์‹คํ–‰ โ€” drivestudio(OmniReยทStreet GSยทPVG) ๋ฅผ ๊ฐ™์€ ์žฅ๋ฉดยท๋ถ„ํ• ๋กœ.
  4. ์‹œ๋ฎฌ๋ ˆ์ดํ„ฐ ์ง€ํ‘œ โ€” HD-Score ํ•ญ๋ชฉ ๊ตฌํ˜„, 6์บ  ๋ฃจํ”„ Hz ์ธก์ •.
  5. ๋ชจ๋“  ์ธ์šฉ ์ˆ˜์น˜๋ฅผ PDF ํ‘œ์™€ ๋Œ€์กฐ (์ด ๋ฌธ์„œ๋Š” ์ดˆ์•ˆ).

์ถœ์ฒ˜ (arXiv)

DrivingGaussian 2312.07920 ยท S-NeRF/NeuRAD 2311.15260 ยท SplatAD 2411.16816 ยท UniSim 2308.01898 ยท EmerNeRF 2311.02077 ยท Street Gaussians 2401.01339 ยท PVG 2311.18561 ยท SยณGaussian 2405.20323 ยท OmniRe 2408.16760 ยท HUGS 2403.12722 ยท DeSiRe-GS 2411.11921 ยท 4DGS 2412.20720 ยท IDSplat 2511.19235 ยท DIAL-GS 2511.06632 ยท ReconDrive 2603.07552 ยท Nighttime PBR-GS 2602.13549 ยท FreeVS 2410.18079 ยท StreetCrafter 2412.13188 ยท DriveDreamer4D 2410.13571 ยท ReconDreamer 2411.19548 ยท ReconDreamer++ 2503.18438 ยท Difix3D+ 2503.01774 ยท OneFixer 2609.32175 ยท Para-Lane 2502.15635 ยท EUVS 2412.05256 ยท HUGSIM 2412.01718 ยท RAD 2502.13144 ยท NeuroNCAP 2404.07762 ยท NAVSIM 2406.15349 / 2506.04218 ยท Bench2Drive 2406.03877 ยท ClimateNeRF 2211.13226 ยท WeatherEdit 2505.20471 ยท Weather-Magician 2505.19919 ยท GSRAIN 2608.02177