This image shows an academic poster presentation at a conference, specifically from the KAIST SGVR Lab. The poster is titled "UFORecon: Generalizable Sparse-View Surface Reconstruction from Arbitrary and Unfavorable Sets." The research was conducted by Youngju Na, Woo Jae Kim, Kyu Beom Han, Suhyeon Ha, and Sung-Eui Yoon. The poster is divided into several sections:

1. **Goal**: Describes the aim of reconstructing surfaces from sparse-view images that include arbitrary and unfavorable view combinations. Illustrative figures and comparative results between different sets and methods are shown.
2. **Challenges**: Lists the main challenges faced in generalizable surface reconstruction, including assumptions on favorably-selected views and difficulties related to training view-combinations. It introduces the View-Combination Score (VC-score) as a metric for evaluating the informativeness of view combinations.
3. **Our Approach**: Details the cross-view correlation-aware volume rendering framework employed by the researchers. Diagrams illustrate the framework components and how it handles unfavorable sets of images.
4. **Core Components**: Breaks down the major elements of their approach, such as the Cascade Cross-View Correlation Framework and methods for random set training.
5. **Quantitative Results**: Provides data and comparisons of the framework against baseline methods across various quantitative metrics.
6. **Qualitative Results**: Includes visual comparisons of reconstructed surfaces from different methods to demonstrate the effectiveness of their approach.
7. **Conclusion**: Summarizes the findings that the UFORecon method generalizes to different view-combinations, providing improved quality in 3D surface reconstructions even with unfavorable sets.
8. **References**: Lists the literature cited by the authors in their research.

The poster is presented under the context of a conference or exhibition, indicating its contribution to the field of computer vision or related areas. A QR code for additional references and possible interactive elements is also included. The setup further includes adjacent posters and a cylindrical object, possibly carrying more information or promotional materials nearby.
Text transcribed from the image:
KNST
SGVR Lab
1. Goal
Generalizable surface reconstruction from sparse-view images with
arbitrary and unfavorable view combinations.
Sparse-view Images
Favorable Set
Unfavorable Set
Favorable Set
Unfavorable Set
UFORecon: Generalizable Sparse-View Surface Reconstruction
from Arbitrary and Unfavorable Sets
Youngju Na Woo Jae Kim Kyu Beom Han Suhyeon Ha Sung-Eui Yoon
3. Our Approach: Cross-view correlation-aware volume rendering framework
FPN
-0
Cross-view Matching
Transformer
W
4.2. Qualitative results
Favorable Set
N Correlation Frustum (Sec. 4.3)
Differentiable Warping
D-DU-d
Concreation
Ret
(144)
Reconstruction
Transformer
(Sec 4.5)
SDF
C
V
V
+
W
DTU [3]
Normal Set
CVPR
SEATTLE, WA JUNE 17-21, 2024
25
25
Unfavorable Set
● 香港科技大學
THE HONG KONG
UNIVERSITY OF SCIENCE
AND TECHNOLOGY
Contril
Volecon
151)
話話話話
A generative NeRF-to-NeRF translation
2
Depth
ValRecon
Outs
VolRecom
Our
Vo
Unfavorable set of N images
(CD: 0.74 (0.59)
0.54
NC2 x 2 Cross-view features.
ŝ= cos(fi, f)
Volume Rendering
(116) (1.00) 0.54
40
671
X-
(C., D.}
task together with a generic solution.
VolRecon
Ours
VolRecon
Ours
2. Challenges
Previous generalizable surface reconstruction methods:
a) Assume favorably-selected view combinations.
b) Easily overfit to training view-combinations.
View-Combination Score (VC-score)
Represents the informativeness of view combinations in reconstruction.
71-1
1
VC =
=1=1+1
s(i,j), s(i,j) =
Σ(G(O₁(P))
P: common 3d track, By: angle between view i and j.
G: piecewise-gaussian function
Chamfer Distance()
VolRecon (Best Set Training)
Ours (Best Set Training)
Ours (Random Set Training)
3.1. Core components
Cascaded Cross-View Correlation Frustum:
Constructs correlation frustums to capture global correlations among
views.
.Geometry-aware Similarity Encoding:
Explicitly encodes pairwise feature similarities as view.
consistent prior.
. Reconstruction Transformers:
Aggregates global and local correlations to estimate geometry accurately.
4.1. Quantitative results
Comparison with baselines on various VC-scores.
(F.F)
Geometry-aware
Similarity Encoding (Sec 4.4).
3.2. Random set training
Randomly selects camera configurations during training to
enhance view-combination robustness.
Effectively integrates with our correlation-aware framework.
T
sampling
BB
888
9
1
BUT 377 MAG
A 3D VAE-GAN framework that can leam
edits corresponding to the a set of input
A contrastive learning framework to di
Superior efficiency, quality, and diver
Blended-MVS [4]
w/ and w/o random set training
V
RTR
Training
M
5. Conclusion
Introduce novel concept in generalizable surface-reconstruction,
view-combination Generalizability.
Effectively combine and validate the efficacy of correlation module
and random set training strategy.
•Achieve SoTA performance on all VC-score levels.
[1] Ren et al. CVPR 2023
Inference
Network pipeline:
Latent distill extract edit codes from the tr
translated NeRF
Optimizie a KL loss to constrain the latere
and Lcontr to optimize the appearance and
Inference: sample a latent vector from Ga
view consistent 3D scene with high quality
Metric: Chamfer Distance (CD) I
Low VC-score leads to
severe degradation!
Set
24
Weken
120
259
1
LOW
142 183 18
134 674
no
12) 127
Favorable
NeTR (2)
LOS
231
144
09
1.42
OAK
1.35
087
1.30 www
1.07 0.77 0.39
VolRecon (1)
Favorable
1.42
105 112
137
(VC 1788)
Our
0.76
205
131
0.82
1.12
05
115
117
LII 87
0.58 0.54
AM
Our
0.71
2.19
134 0.47
1.16
125
117
150
0.57 SL
289 249
265
112
264 110
217
The
Normal
RTR (2)
206
2.31
231
1.75
211
1.09
134 1.35
"
LM
205
174
La
153
ReTR* (2)
5%
(VC 192)
Ours
2.59 131
12
104
ADAY
C ALADD
137
1.16
095
M
100 LO
12
Ours
Random set training Metric: Chamfer Distance (CD)
Unfavorable
Method
3.18
2.94
VolRecon [1] 2.74 (+1.32.92%) 3.88 (+0.7, 22)
ReTR (2)
1.62 (+0.45,38%) 2.88 (-0.06.20)
099
6. References
156
Oury
221
142
123 124
8.72
037
BAS
La
Ours
1.01 (+0.02.2%) 128 (-0.28, 18%)
(2) Liang et al. NeuriPS 2023
164 $26
46)
243
LS
"THIE 241 236
249 334 335 1.44 360 33
(3) Aanes et al. UCV 2016
10%
102
10:
Unfavorable
RTR (2)
3.90
17
422
222
2.93
1 231
224
236 236
192 163 283
2H
First row: best set training.
(VC 57)
Oury
199 2.23
LAS
122
1.92 130
1.25
120 1.29
13
(4) Yao et al. CVPR 2020
View-Combination Score
Our
131
LAS
FY 1.37 0.99
M
LAS
W
LM
Second row: random set training.
200