Unsigned Orthogonal Distance Fields: An Accurate Neural Implicit Representation for Diverse 3D Shapes

CVPR 2024

Yujie Lu1, Long Wan1, Nayu Ding1, Yulong Wang1, Shuhan Shen2,4, Shen Cai1*, Lin Gao3,4*

1Visual and Geometric Perception Lab, Donghua University    2Institute of Automation, Chinese Academy of Sciences    3Institute of Computing Technology, Chinese Academy of Sciences    4University of Chinese Academy of Sciences.

Paper

Code

Video

Abstract

Description in one sentence: Different from SDF and UDF, each point in UODFs defines the distances along the three orthogonal directions (LR, FB, UD in the figure) and directly infers the nearest surface points (simulating the characteristics of the laser), avoiding the introduction of interpolation errors and improving the reconstruction accuracy.




Reconstruction of Watertight Shapes




Reconstruction of Non-watertight Shapes



Reconstruction of Complex Shapes



Reconstruction at Multi-resolution Grids




Reconstruction Results of Grid Edge Points (GEP)





Citation

@inproceedings{UODFs,
    title={Unsigned Orthogonal Distance Fields: An Accurate Neural Implicit Representation for Diverse 3D Shapes},
    author={Lu, Yujie and Wan, Long and Ding, Nayu and Wang, Yulong and Shen, Shuhan and Cai, Shen and Gao, Lin},
    booktitle={IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year={2024}
}