Deformed Implicit Field: Modeling 3D Shapes with Learned Dense Correspondence

This is an official pytorch implementation of the following paper:

Y. Deng, J. Yang, and X. Tong, Deformed Implicit Field: Modeling 3D Shapes with Learned Dense Correspondence, IEEE Computer Vision and Pattern Recognition (CVPR), 2021.

Paper | Slides

Abstract: We propose a novel Deformed Implicit Field (DIF) representation for modeling 3D shapes of a category and generating dense correspondences among shapes. With DIF, a 3D shape is represented by a template implicit field shared across the category, together with a 3D deformation field and a correction field dedicated for each shape instance. Shape correspondences can be easily established using their deformation fields. Our neural network, dubbed DIF-Net, jointly learns a shape latent space and these fields for 3D objects belonging to a category without using any correspondence or part label. The learned DIF-Net can also provides reliable correspondence uncertainty measurement reflecting shape structure discrepancy. Experiments show that DIF-Net not only produces high-fidelity 3D shapes but also builds high-quality dense correspondences across different shapes. We also demonstrate several applications such as texture transfer and shape editing, where our method achieves compelling results that cannot be achieved by previous methods.

Features

● Surface reconstruction

We achieve comparable results on surface reconstruction task with other implicit-based shape representation.

● Dense correspondence reasoning

Our model gives reasonable dense correspondence between shapes in a category.

● Awareness of structure discrepancy

Our model predicts correspondence uncertainty between shapes in a category, which depicts structure discrepancies.

Installation

1. Clone the repository and set up a conda environment with all dependencies as follows:

git clone https://github.com/microsoft/DIF-Net.git --recursive
cd DIF-Net
conda env create -f environment.yml
source activate dif

2. Install torchmeta. Before installation, comment out line 3 in pytorch-meta/torchmeta/datasets/utils.py, otherwise the library cannot be imported correctly. Then, run the following script:

cd pytorch-meta
python setup.py install

Generating shapes with pre-trained models

1. Download the pre-trained models from this link and organize the directory structure as follows:

DIF-Net
│
└─── models
    │
    └─── car
    │   │
    |   └─── checkpoints
    |       |
    |       └─── *.pth
    │
    └─── plane
    │   │
    |   └─── checkpoints
    |       |
    |       └─── *.pth
    ...

2. Run the following script to generate 3D shapes using a pre-trained model:

# generate 3D shapes of certain subjects in certain category
python generate.py --config=configs/generate/<category>.yml --subject_idx=0,1,2

The script should generate meshes with color-coded template coordinates (in ply format) into ./recon subfolder. The color of a surface point records the 3D location of its corresponding point in the template space, which indicates dense correspondence information. We recommand using MeshLab to visualize the meshes.

Training a model from scratch

Data preparation

We provide our pre-processed evaluation data from ShapeNet-v2 as an example. Data can be download from this link (four categories, and 100 shapes for each category respectively. 7 GB in total). The data contains surface points along with normals, and randomly sampled free space points with their SDF values. The data should be organized as the following structure:

DIF-Net
│
└─── datasets
    │
    └─── car
    │   │
    |   └─── surface_pts_n_normal
    |   |   |
    |   |   └─── *.mat
    │   |
    |   └─── free_space_pts
    |       |
    |       └─── *.mat    
    |
    └─── plane
    │   │
    |   └─── surface_pts_n_normal
    |   |   |
    |   |   └─── *.mat
    │   |
    |   └─── free_space_pts
    |       |
    |       └─── *.mat    
    ...

To generate the whole training set, we follow mesh_to_sdf provided by marian42 to extract surface points and normals, as well as calculate SDF values for ShapeNet meshes. Please follow the instruction of the repository to install it.

Training networks

Run the following script to train a network from scratch using the pre-processed data:

# train dif-net of certain category
python train.py --config=configs/train/<category>.yml

By default, we train the network with a batchsize of 256 for 60 epochs on 8 Tesla V100 GPUs, which takes around 4 hours. Please adjust the batchsize according to your own configuration.

Evaluation

To evaluate the trained models, run the following script:

# evaluate dif-net of certain category
python evaluate.py --config=configs/eval/<category>.yml

The script will first embed test shapes of certain category into DIF-Net latent space, then calculate chamfer distance between embedded shapes and ground truth point clouds. We use Pytorch3D for chamfer distance calculation. Please follow the instruction of the repository to install it.

Contact

If you have any questions, please contact Yu Deng (dengyu2008@hotmail.com) and Jiaolong Yang (jiaoyan@microsoft.com)

License

Copyright © Microsoft Corporation.

Licensed under the MIT license.

Citation

Please cite the following paper if this work helps your research:

@inproceedings{deng2021deformed,
	title={Deformed Implicit Field: Modeling 3D Shapes with Learned Dense Correspondence},
	author={Yu Deng and Jiaolong Yang and Xin Tong},
    booktitle={IEEE Computer Vision and Pattern Recognition},
    year={2021}
}

Acknowledgement

This implementation takes SIREN as a reference. We thank the authors for their excellent work.