3D_Semantic_Subspace_Traverser (ICCV 2023)

3D Semantic Subspace Traverser: Empowering 3D Generative Model with Shape Editing Capability

Ruowei Wang, Yu Liu, Pei Su, Jianwei Zhang, Qijun Zhao

This is the official repository for the paper "3D Semantic Subspace Traverser: Empowering 3D Generative Model with Shape Editing Capability".

Overview

Abstract:
Shape generation is the practice of producing 3D shapes as various representations for 3D content creation. Previous studies on 3D shape generation have focused on shape quality and structure, without or less considering the importance of semantic information. Consequently, such generative models often fail to preserve the semantic consistency of shape structure or enable manipulation of the semantic attributes of shapes during generation. In this paper, we proposed a novel semantic generative model named 3D Semantic Subspace Traverser that utilizes semantic attributes for category-specific 3D shape generation and editing. Our method utilizes implicit functions as the 3D shape representation and combines a novel latent-space GAN with a linear subspace model to discover semantic dimensions in the local latent space of 3D shapes. Each dimension of the subspace corresponds to a particular semantic attribute, and we can edit the attributes of generated shapes by traversing the coefficients of those dimensions. Experimental results demonstrate that our method can produce plausible shapes with complex structures and enable the editing of semantic attributes.

Citation

If you make use of our work, please cite our paper:

@inproceedings{ruowei2023traverser,
  title={3D Semantic Subspace Traverser: Empowering 3D Generative Model with Shape Editing Capability},
  author={Wang, Ruowei and Liu, Yu and Su, Pei and Zhang, Jianwei and Zhao, Qijun},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  year={2023}
}

Requirements

The code is developed with

• Python 3.7
• Pytorch 1.8.0
• Cuda 11.4

You can also create an environment using the setting file by conda env create -f environment.yml.

Data Preprocessing

Our data preprocessing code is based on IF-NET and DISN. Please refer them if you this code. And if you have any questions, you can read the original code of IF-NET and DISN.

The following data preprocessing codes are customized to the dataset ShapeNetCore.v1, but you can also apply them to any *.obj files after editing.

First, install the needed libraries with:

cd data_processing/libmesh/
python setup.py build_ext --inplace
cd ../libvoxelize/
python setup.py build_ext --inplace
cd ../..

Second, please set the following files executable:

data_processing/bin/PreprocessMesh
data_processing/bin/SampleVisibleMeshSurface
data_processing/isosurface/computeDistanceField
data_processing/isosurface/computeMarchingCubes
data_processing/isosurface/displayDistanceField

Third, we offer some data samples in sample_data folder. You can process the data with:

python data_processing/1_create_isosurf.py --input_path=../sample_data/ShapeNetCore.v1 --thread_num=8 --category=chair
python data_processing/2_convert_to_scaled_off.py  --processed_data_dir=../sample_data/ShapeNetCore.v1_processed_tmp
python data_processing/3_voxelize.py --processed_data_dir=../sample_data/ShapeNetCore.v1_processed_tmp -res=64
python data_processing/4_boundary_sampling.py --processed_data_dir=../sample_data/ShapeNetCore.v1_processed_tmp

we save the processed data in ../sample_data/ShapeNetCore.v1_processed_tmp, which is defined in a dictionary named data in the data_processing/1_create_isosurf.py.

Training Code

1. we train the VAE with:

python train_VAE.py --path=<str>

• path is the path of the processed dataset. ./sample_data/ShapeNetCore.v1_processed_tmp by default.

All training results of VAE are saved in ./run_CubeCodeCraft/****-AE-batch*-steps*.

2. we can use the trained VAE to generate the latent code for training the GAN:

python VAE_produce_data.py path_of_VAE_ckpt

• path_of_VAE_ckpt is the path of a trained VAE ckpt produced by step 1.

It produces shape codes in ./sample_data/ShapeNetCore.v1_processed_tmp_encoded by default.

3. we use these shape codes to train the GAN:

python train_GAN.py \
--path=<str> \
--ae_ckpt_path=<str>

• path is the path of encoded data produced by step 2. ./sample_data/ShapeNetCore.v1_processed_tmp_encoded by default.
• ae_ckpt_path is the path of a trained VAE ckpt produced by step 1.

All training results of GAN are saved in ./run_CubeCodeCraft/****-GAN-batch*-steps*.

Pre-training Models

Baidu Netdisk: https://pan.baidu.com/s/1BAytVXCq5D3zOdtdkF_inA?pwd=68jr, (password=68jr)

Dropbox: https://www.dropbox.com/scl/fo/b1yvcsk8yx1rl48gyhb3c/h?rlkey=4ruwe4yzx43q4vpaynhvw4at1&dl=0

Inference

generation

python GAN_generate_data.py path_of_GAN_ckpt --ae_ckpt_path=<str> --num_generated=<int>

• path_of_GAN_ckpt is the path of a trained GAN ckpt.
• ae_ckpt_path is the path of a trained VAE ckpt.
• num_generated is the number of generated results.

Generated results are saved in a folder named as generate_*_r=*_threshold=*_***_num* nearby path_of_GAN_ckpt.

shape manipulation (or shape exploration) by the local linear subspace models.

python GAN_generate_data_Eigen_newRender.py path_of_GAN_ckpt --ae_ckpt_path=<str> \
--traverse_range=<float> \
--intermediate_points=<int>
--tvs_seed <int> <int> ……

• path_of_GAN_ckpt is the path of a trained GAN ckpt.
• ae_ckpt_path is the path of a trained VAE ckpt.
• traverse_range is the number of generated results.
• intermediate_points is the number of intermediate points during traversing.
• tvs_seed are seeds to generate shapes for manipulation.

Generated results are saved in a folder named as generateEigen_*_r=*_threshold=*_*** nearby path_of_GAN_ckpt.

Acknowledgements

our code is heavily built upon IF-NET and EigenGAN. If you find my code useful, please also consider to cite those papers.

Contact

Besides the email address in the paper, feel free to send an email to 772438854@qq.com.

LICENSE

LICENSE.

TODO

• overview
• requirements
• data preprocessing
• training code
• pre-training models
• inference code
• acknowledgements
• contact
• license