Structure-Aware 3D Shape Synthesis from Single-View Images [Paper]

TensorFlow implementation for “Structure-Aware 3D Shape Synthesis from Single-View Images”.

Dependencies

Python>=3.5, TensorFlow>=1.4, numpy, scipy, matplotlib

Dataset

We used ShapeNet dataset Download dataset, password: se16. We provided the training set in this link, if you need the full of dataset, please contact me.

Testing

Results of our Download pretrained models, password: 3j2z. We provided the pretrained model in $—MODE leftside_to_rightside$, if you need the full of pretrained model, please contact me.

python main.py --phase test --mode $MODE$ --test_data_path $DATA_DIR$ --pretrained_model_path $PRETRAINED_MODEL$

• MODE: src viewpoint to dst viewpoint, such as leftside_to_rightside
• DATA_DIR: path to test data
• PRETRAINED_MODEL: path of the pretrained model

Then, you can use 3D-R2N2 test based on the synthesis images.

Training

python main.py --phase train --data_path $DATA_DIR$ --mode $MODE$

• DATA_DIR: path to train data
• MODE: src viewpoint to dst viewpoint, such as leftside_to_rightside

Then, you can use 3D-R2N2 train based on the synthesis images.

Citation

If you use any part of this code in your research, please cite our paper:

@inproceedings{hu2020structure,
  title={Structure-Aware 3D Shape Synthesis from Single-View Images},
  author={Xuyang Hu, Fan Zhu, Li Liu, Jin Xie, Jun Tang, Nian Wang, Fumin Shen, and Ling Shao},
  booktitle={Proceedings of the 29th British Machine Vision Conference},
  year={2018}
}

Acknowledge

The code is built on CycleGAN (TensorFlow) and 3D-R2N2. We thank the authors for sharing the codes.