MingyuY / Iterative-view-synthesis

ID-Unet: Iterative-view-synthesis(CVPR2021 Oral)

Tensorflow implementation of ID-Unet: Iterative Soft and Hard Deformation for View Synthesis.

Overview architecture

The network architecture

Experiment Results

• chair

• MultiPIE

• Flow

Preparation

• Prerequisites
  • Tensorflow
  • Python 2.x with matplotlib, numpy and scipy
• Dataset
  • 3D chair
  • MultiPIE
  • 300w-LP
  • Images should be placed in ./datasets/
• Download model
  • Tool model
    • model should be placed in ./models/

Quick Start

Exemplar commands are listed here for a quick start.

dataset

• prepare dataset

  python datasets/creat_txt.py --path_MultiPIE 'Path to MultiPIE Dataset' --path_chair 'Path to chair Dataset' --path_300w_LP 'Path to 300w-LP Dataset'
  
  shuf datasets/multiPIE_train_paired.txt -o  datasets/multiPIE_train_paired_shuf.txt
  
  python datasets/creat_tf.py --path_MultiPIE 'Path to MultiPIE Dataset' --path_chair 'Path to chair Dataset' --path_300w_LP 'Path to 300w-LP Dataset'
  

Training

• To train with size of 128 X 128

  python MultiPIE.py --mode training
  
  python chair.py --mode training

Testing

• Example of test

  python  MultiPIE.py --mode test --batch_size 1 --model_path 'Path to Training Model'
  
  python  chair.py --mode test --batch_size 1 --model_path 'Path to Training Model'

Citation

If this work is useful for your research, please consider citing:

@inproceedings{yin2021id, title={ID-Unet: Iterative Soft and Hard Deformation for View Synthesis}, author={Yin, Mingyu and Sun, Li and Li, Qingli}, booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, pages={7220--7229}, year={2021} }