This is a PaddlePaddle2.0 implementation of the paper 《Pix2Vox: Context-aware 3D Reconstruction from Single and Multi-view Images》

See the official repo in Pytorch, as well as overview of the method.

AI Studio Notebook.

differences

1. In order to maintain the same network structure as the original paper, this experiment used vgg16_bn without pre-trained(PaddlePaddle has vgg16 model pre-trained on ImageNet, but no pre-trained vgg16_bn model. ).

This may have affected the experimental results.

Use the same dataset as mentioned in the official repo.

--ShapeNet rendered images http://cvgl.stanford.edu/data2/ShapeNetRendering.tgz

--ShapeNet voxelized models http://cvgl.stanford.edu/data2/ShapeNetRendering.tgz

--Pix3D images & voxelized models: http://pix3d.csail.mit.edu/data/pix3d.zip

The dataset is already mounted in this notebook.

!unzip -oq data/data67155/dataset.zip

%cd work/Pix2Vox-F/
!pip install -r requirements.txt

%cd work/Pix2Vox-F/

# train
!python runner.py

# test
!python3 runner.py --test --weights=/path/to/best_checkpoint

%cd work/Pix2Vox-A/

# train
!python runner.py

# test
!python3 runner.py --test --weights=/path/to/best_checkpoint