A conference attendee engages with a highlighted poster presentation titled "Sat2Scene: 3D Urban Scene Generation from Satellite Images with Diffusion" at a technology and research event. The poster, featuring affiliations with institutions such as ETH Zurich, UTokyo, Microsoft, and the University of Amsterdam, was authored by Zuoyue Li, Zhenqiang Li, Zhaopeng Cui, Marc Pollefeys, and Martin R. Oswald. The detailed content of the poster includes sections on the task, experiments, baseline comparisons, ablation studies, and model generalization, reflecting a comprehensive study on generating 3D urban scenes from satellite imagery using diffusion models. A nearby attendee wears a lanyard with the phrase "Sharpen Your AI with Better Data," indicating the event's focus on advanced technology and data. Another participant, holding a laptop, appears to be engaged in further discussions or demonstrations related to the presented research, accentuating the collaborative and informative atmosphere of the conference.
Text transcribed from the image:
Highlight
ETH Zürich
Sat2Scene: 3D Urban Scene Generation from Satellite Images with Diffusion
UTokyo |浙江大学 Microsoft
ntroduction
fask Generate 3D urban scene on a give
arbitrary 2D views with robust co
UNIVERSITY
OF AMSTERDAM
Zuoyue Li Zhenqiang Li Zhaopeng Cui Marc Pollefeys Martin R. Oswald
render
4 Experiment
CVPR
SEATTLE, WA JUNE 17-21, 2024
Baseline comparison Method/Metric | FVD↓
HoliCity dataset
GT geometry
Sat2 Vid
37.06
KVD100+
4.03 0.05
InfiniCity
Various metrics
solds
ectory
MVDiffusion
Ours
22.79
20.30
2.36003
1.90±0.03
KID100+ PSNR
FID
25.25
137.84 13.760.10
108.47 8.40±0.10
50.78
4.14007
71.98
5.91006
SSIM LPIPS User study
0.741
2.92%
0.252
Ablation study
0.593
17.56
0.956
31.54
0.259
0.237 81.46%)
15.62%
w/o point resampling
w/o point aggregation
w/o depth supervision
w/o & w/ point resampling point weights
Exemplary
scene used
els
Sharpen Your Al with Better Data Shar
e
ining
Is
Sat
GT
Ours
GT
Sat2Vid
MVDiffusion
for training
Variant/Metric FID↓
KID 100 Dep. RMSE
w/o pt-rsmp
w/o pt-aggr
131.38
85.58
12.66012
7.79008
3.22
w/o dep-sup
80.40
7.2200
3.441
Ours
71.98
5.91006
3.07
w/o pt-aggr
w/o dep-sup
Ours
w/o pt-rsmp
w/o pt-aggr
w/o dep-sup
Ours
Sat2Vid
MVDiffusion
w/o pt-rsmp
Ours
Model generalization OmniCity dataset, long-seq generation on predicted geometry
Grou Mu
Ground-view
Ours
Bird-view
C
.
.
In
Sat Scene: 30 Urban Scene Generation
Pho
Larg
Future
3D spar
n Satellite ages with Diffusion
Advanced
Conditional gen
OPK24
Depths
rXiv
228