Yongdong Huang, Yuanzhan Li, Xulong Cao, Siyu Zhang, Shen Cai*, Ting Lu, Jie Wang, and Yuqi Liu. An Efficient End-to-End 3D Model Reconstruction based on Neural Architecture Search. Accepted by ICPR2022 (oral presentation).
We complete an end-to-end neural network for 3D model reconstruction task by classifying binary voxels and utilizing the technology of neural architecture search (NAS). Compared to other signed distance field (SDF) prediction or binary classification methods, our method achieves significantly higher reconstruction accuracy using fewer network parameters.
ONet , NI and NGLOD are the methods we compared in our paper.
[ONet] L. Mescheder, M. Oechsle, M. Niemeyer, S. Nowozin, and A. Geiger, “Occupancy networks: Learning 3d reconstruction in function space,”, in CVPR, 2019.
[NI] Thomas Davies, Derek Nowrouzezahrai, and Alec Jacobson, “On the effectiveness ofweight-encoded neural implicit 3d shapes,” arXiv:2009.09808, 2020.
[NGLOD] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler, “Neural geometric level of detail: real-time rendering with implicit 3d shapes,” in CVPR, 2021.
Models and reconstructions of the 3D objects presented in our paper are in the RESULTS folder. You can directly run eval.py to get the reconstruction results, or download Meshlab to open the reconstruction results provided by us. Note that when using MeshLab to view the reconstruction results we provide, select X Y Z for Point format and SPACE for Separator.
We use Shapenet and Thingi10k datasets, both of which are available from their official websites. Thingi32 is composed of 32 simple shapes in Thingi10K. ShapeNet150 contains 150 shapes in the ShapeNet dataset.
cd ./
bash train.sh
python eval.py
We verified that it worked on ubuntu18.04 & cuda10.2.
python 3.6
tensorflow 2.0
This project is licensed under the terms of the LGPL License (see LICENSE
for details).