The image shows a detailed academic poster presented at the CVPR (Computer Vision and Pattern Recognition) conference held in Seattle, WA from June 18-22, 2023. The poster is titled "2S-UDF: A Novel Two-stage UDF Learning Method for Robust Non-watertight Model Reconstruction from Multi-view Images" by authors Junkai Deng, Fei Hou, Xuhui Chen, Wencheng Wang, and Ying He, affiliated with Nanyang Technological University and Chinese Academy of Sciences.

The poster introduces a method for reconstructing non-watertight 3D models from multiple view images, emphasizing the Objective and detailing the methodology across Key Stages. Stage 1 approximates UDF learning with volume rendering concepts, while Stage 2 refines UDF through a direct weight mapping approach.

Different sections include comparisons with benchmarks such as DTU Benchmark and Other Benchmarks, illustrating the efficiency of their method against existing ones like NeuralUDF, NeuDF, and NeAT. The pipeline section outlines the steps involved, from volume rendering to loss function. Figures and illustrations provide visual comparisons of model reconstructions, and detailed reference sections cite critical background literature. This visually informative poster stands as a substantial contribution to advancements in 3D model reconstruction technology.
Text transcribed from the image:
ISCAS
NANYANG
TECHNOLOGICAL
UNIVERSITY
④ 中国科学院大学
University of Chinese Academy of Sciences
2S-UDF: A Novel Two-stage UDF Learning Method for Robust Non-watertight
Model Reconstruction from Multi-view Images
Junkai Deng
Fei Hou
Xuhui Chen
-KEY STAGES
Wencheng Wang
Ying He
OBJECTIVE
Our objective is to reconstruct non-watertight models
from multi-view images.
Stage 1: Approximate UDF Learning-Volume Rendering
We use the NeuS [4]-like volume rendering pipeline.
We map the distance value f(r(t)) to the volume density σ(r(t)) by
RELATED
NeuralUDF [1] Introduces indicator function increasing rendering
complexity and volatileness.
NeUDF [2] Requires numerical approximation of unbounded integral.
NeAT [3] Requires object masks for training.
PIPELINE
Dist.
Pred.
MLP
Softplus
Rendering
Color
Pred.
MLP
Param.
BG.
Pred.
MLP
"Key Stages"
1&2
SULAR
Composition
Background
Rendered
Image
o(r(t))
se-f(r())
1+e¯f(r(t))
C
Stable: No unbounded integral: Opaque: X (almost) Unbiased: x (almost)
Stage 2: UDF Refinement-Direct Weight Mapping
We directly map the UDF value to the color weight.
Opaque: ✓
A UDF maxima, accumulated weight below
threshold. Not a truncation point.
B On the surface.
C Ray truncation point.
D On the surface. Ray already truncated-no
color contribution.
Unbiased:
DEEPFASHION3D BENCHMARK-
GT
Ours
Loss
NeuralUDF [1]
NeUDF [2]
I GT
Image
GT
GT
DTU BENCHMARK
Ours
CVPR
SEATTLE, WA JUNE 17-21, 2024
NeUDF [2]
11
OTHER BENCHMARK
Ours
NeUDF [2] Neural UDF [1] NeAT [3]
REFERENCES
[1] Long et al. NeuralUDF: Learning Unsigned Distance Fields for Multi-View Reconstruction of
Surfaces with Arbitrary Topologies. CVPR 2023.
[2] Liu et al. NeUDF: Leaning Neural Unsigned Distance Fields with Volume Rendering. CVPR
2023.
[3] Meng et al. NeAT: Learning Neural Implicit Surfaces with Arbitrary Topologies from Multi-
View Images. CVPR 2023.
[4] Wang et al. NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-view
Reconstruction. NeuriPS 2021.
24