KAIST_CS492_DeepImplicitTemplates

KAIST CS492(A) Assignment

This repository is an implementation for Deep Implicit Templates. And the paper is Deep Implicit Templates.

Requirements

• Ubuntu(v18.04)
• numpy // 1.22.3
• Pytorch // 1.9.0+cu111
• sklearn // 0.24.1
• ninja // 1.10.2.3
• pathos // 0.2.8
• trimesh // 3.10.2
• skimage // 0.19.2
• pyrender // 0.1.45
• scipy // 1.8.0
• plyfile
• easydict
• tqdm
• os

For Data Preprocessing

• pyopengl // 3.1.0
• pygame // 2.1.2
• mesh_to_sdf // 0.0.14
• opencv-python // 4.5.5.64

Data Preprocessing

Our preprocessing code is highly based on mesh2sdf project and ShapeGAN. It takes about 1-2 days to preprocess the all SDF files. So we prepared preprocessed dataset(04256520.egg) in Gdrive Link. You can easily download and try it.

python prepare_data_dir-pn.py

Implementataion

A benchmark dataset of our project is ShapeNetCore.v2. And you can download preprocessed datasets(SDF files) and pre-trained model from this Gdrive Link.

git clone https://github.com/Goya-Studio/KAIST_CS492_DeepImplicitTemplates.git
cd KAIST_CS492_DeepImplicitTemplates

1. Deep Implicit Templates.ipynb : Implement and train DIT model.
2. Reconstruction.ipynb : Reconstruct a shape using trained model.(latent code train again.)
3. Evaluation.ipynb : Evaluate the result by EMD, CD metrics.

• You need to download ShapeNetCore.v2 for evaluation.

• If you download the ShapeNet data, please put the data in gt folder.

  

Acknowledgements

This repository is based on DeepSDF and Deep Implicit Templates. And we use the mesh2sdf project to preprocess the data. Thank you all!

Reference Project

• DeepSDF
• Deep Implicit Templates
• mesh2sdf
• ShapeGAN
• ShapeNetCore.v2

[NOTE]

• If you have a problem about 'pip install easydict', please git clone this project.

Contact

• Yongmin: [email protected]
• Olivia: [email protected]