第９回全日本コンピュータビジョン勉強会 ICCV2021論文読み回発表資料 GNeRF: GAN-based Neural Radiance Field without Posed Camera
GNeRF: GAN-based Neural RadianceField without Posed Camera@peisuke9
View Slide
ABEJATwitter @peisukeGithub https://github.com/peisukeQiita https://qiita.com/peisukeSlideShare https://www.slideshare.net/FujimotoKeisuke
•• GNeRF: GAN-based Neural Radiance Field without Posed Camera•• NeRF••• GAN Radiance Field• RF
NeRF• / 5•
••(x,y,z)T(φ1,θ1)T(φ2,θ2)T(x,y,z) φ2,θ2(x,y,z) φ1,θ1
NeRF•••
NeRF• NeRF in the Wild• NeRF• iNeRF•iNeRFNeRF in the Wild BP
条件 iNeRF NeRF-- IDR GNeRF(提案⼿法)事前学習 必要 不要 不要 不要粗いカメラ座標 必要 必要 必要 不要前⽅のみの制約 不要 必要 不要 不要マスク 不要 不要 必要 不要••
••••• GAN Radiance Field• RF
• Phase A Phase B
RF• DE 3
GAN Real•• Discriminator Real• Dynamic Patch Sampling /16x16• Dynamic Patch Sampling
GAN Fake•• NeRF16x16• G
• 6𝑹 =𝑟!!𝑟!"𝑟!#𝑟"!𝑟""𝑟"#𝑟#!𝑟#"𝑟##𝒓 = 𝑟!!𝑟!"𝑟!#𝑟"!𝑟""𝑟"#$• SO(3)NN• On the Continuity of Rotation Representations in Neural Networks
• Inversion network• Static Patch Sampling• E• Static Patch Sampling
GGD G
•• Inversion network• refinement
•Refinement
••••• Photometric loss
•• Generator• NeRF• Discriminator• Instance Normalization Spectral Normalization• Inversion network• Vision Transformer•• A• A B 1 iteration• B
•• COLMAP NeRF C+n• Refine C+r• Ground Truth NeRF G+n•• Synthetic-NeRF image size: 400x400 DTU 500x400• 100•• Peak Signal-to-Noise Ratio SSIM LPIPS• /• Single TAITAN GPU 30
COLMAP•
• COLMAPDrums
難しいシーン特徴の少ないシーン
IDR• 3DIDR• IDR• PNSR•
• COLMAP• Drums Lego COLMAP
Ablation Study••GAN InvNet. PhaseB•• A B A ABABABAB B
••• COLMAP••
• NeRF•• GAN RF• RF••• COLMAP
peisuke
autifyhq
ytaisei
pauli
soracom
kochizufan
gl_tsukasa
ama_ch
saboyutaka
tsujiba
line_developers
mitsuakitohei
soudai
moore
danielanewman
andyhume
kevinliebholz
gr2m
mongodb
zakiyama
reverentgeek
addyosmani
hannesfritz
philnash
robhawkes