Awesome

GANSeg: Learning to Segment by Unsupervised Hierarchical Image Generation (CVPR 2022)

GANSeg: Learning to Segment by Unsupervised Hierarchical Image Generation <br> Xingzhe He, Bastian Wandt, and Helge Rhodin <br> IEEE/CVF Conference on Computer Vision and Pattern Recognition CVPR 2022

[Paper]

Setup

Setup environment

conda create -n ganseg python=3.8
conda activate ganseg
pip install -r requirements.txt

Download datasets

The CelebA-in-the-wild, Taichi, CUB and Flower can be found on their websites. We provide the pre-processing code for CelebA-in-the-wild, CUB and Flower to make them h5 files. Taichi can be used directly.

Download pre-trained models

The pre-trained models (GAN and Segmenter) can be downloaded from Google Drive.

Testing

Segmentation

You can use gen_mask.py to generate the segmentation masks.

python gen_mask.py --segmenter_log log/seg_celeba_wild_k8 --test_class_name mafl_wild_test --data_root data/celeba_wild --save_root saved_mask/celeba_wild_k8 --checkpoint 10

where,

• --segmenter_log specifies the log folder of the segmentation network,
• --test_class_name specifices the particular dataset to test,
• --data_root specifies the location of the dataset (the folder containing the h5 file),
• --save_root defines the location of the saved images, and
• --checkpoint specifies the index of the checkpoint.

Therefore, the above command will generate masks on the CelebA-in-the-wild.

You can also quantitatively test the segmentation.

python test_seg.py --segmenter_log log/seg_celeba_wild_k8 --test_class_name mafl_wild_test --data_root data/celeba_wild --checkpoint 10

GAN

You can use gen_img.py to generate images with our GAN.

python gen_img.py --generator_log log/gan_celeba_wild_k8 --save_root saved_image/celeba_wild_k8 --checkpoint 30

Training

GAN

To train our GAN on CelebA-in-the-wild, run

python train_gan.py --class_name celeba_wild --data_root data/celeba_wild --n_keypoints 8

The trained weights and log can be found in logs/gan_celeba_wild_k8.

We also provide a custom choice for class_name. You can specify data_root to your own image folder to train our GAN on your own images.

To finetune the learned model, run

python train_gan.py --class_name celeba_wild --data_root data/celeba_wild --n_keypoints 8 --checkpoint [the epoch index to start]

The example parameters can be found in log.

Segmentation

To train the segmenter on the pre-trained GAN (CelebA-in-the-wild), run

python train_seg.py --generator_log log/gan_celeba_wild_k8 --data_root data/celeba_wild --checkpoint 30

where,

• --generator_log specifies the generator log folder (used to generate image-mask pairs),
• --data_root specifies the location of the dataset, and
• --checkpoint specifies the index of the checkpoint of the GAN.

Citation

@inproceedings{he2022ganseg,
  title={GANSeg: Learning to Segment by Unsupervised Hierarchical Image Generation},
  author={He, Xingzhe and Wandt, Bastian and Rhodin, Helge},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={1225--1235},
  year={2022}
}

Acknowledgements

The code is built upon LatentKeypointGAN.