Incremental Transformer Structure Enhanced Image Inpainting with Masking Positional Encoding

by Qiaole Dong*, Chenjie Cao*, Yanwei Fu

Paper and Supplemental Material (arXiv)

Our project page is available at https://dqiaole.github.io/ZITS_inpainting/.

🔥🔥🔥 News: Our Extended version ZITS++ has been accepted by TPAMI, codes and dataset have been released in here.

Pipeline

The overview of our ZITS. At first, the TSR model is used to restore structures with low resolutions. Then the simple CNN based upsampler is leveraged to upsample edge and line maps. Moreover, the upsampled sketch space is encoded and added to the FTR through ZeroRA to restore the textures.

TO DO

• Releasing inference codes.
• Releasing pre-trained model.
• Releasing training codes.

Preparation

1. Preparing the environment:

   as there are some bugs when using GP loss with DDP (link), we strongly recommend installing Apex without CUDA extensions via torch1.9.0 for the multi-gpu training

   conda create -n train_env python=3.6
   conda activate train_env
   pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html
   pip install -r requirement.txt
   git clone https://github.com/NVIDIA/apex
   cd apex
   pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" ./
   

2. For training, MST provide irregular and segmentation masks (download) with different masking rates. And you should define the mask file list before the training as in MST.

   The training masks we used are contained in coco_mask_list.txt and irregular_mask_list.txt, besides test_mask.zip includes 1000 test masks.

3. Download the pretrained masked wireframe detection model to the './ckpt' fold: LSM-HAWP (MST ICCV2021 retrained from HAWP CVPR2020).

4. Prepare the wireframes:

   Update: No need prepare another environment anymore, just extract wireframes with following code

   conda activate train_env
   python lsm_hawp_inference.py --ckpt_path <best_lsm_hawp.pth> --input_path <input image path> --output_path <output image path> --gpu_ids '0'
   

5. If you need to train the model, please download the pretrained models for perceptual loss, provided by LaMa:

   mkdir -p ade20k/ade20k-resnet50dilated-ppm_deepsup/
   wget -P ade20k/ade20k-resnet50dilated-ppm_deepsup/ http://sceneparsing.csail.mit.edu/model/pytorch/ade20k-resnet50dilated-ppm_deepsup/encoder_epoch_20.pth
   

6. Indoor Dataset and Test set of Places2 (Optional)

   To download the full Indoor dataset: BaiduDrive, passward:hfok; Google drive (link).

   The training and validation split of Indoor can be find on indoor_train_list.txt and indoor_val_list.txt.

   The test set of our Places2 can be find on places2_test_list.txt.

Eval

Download pretrained models on Places2 here.

Link for BaiduDrive, password:qnm5

Batch Test

For batch test, you need to complete steps 3 and 4 above.

Put the pretrained models to the './ckpt' fold. Then modify the config file according to you image, mask and wireframes path.

Test on 256 images:

conda activate train_env
python FTR_inference.py --path ./ckpt/zits_places2 --config_file ./config_list/config_ZITS_places2.yml --GPU_ids '0'

Test on 512 images:

conda activate train_env
python FTR_inference.py --path ./ckpt/zits_places2_hr --config_file ./config_list/config_ZITS_HR_places2.yml --GPU_ids '0'

Single Image Test

This code only supports squared images (or they will be center cropped).

conda activate train_env
python single_image_test.py --path <ckpt_path> --config_file <config_path> \
 --GPU_ids '0' --img_path ./image.png --mask_path ./mask.png --save_path ./

Training

⚠️ Warning: The training codes is not fully tested yet after refactoring

Training TSR

python TSR_train.py --name places2_continous_edgeline --data_path [training_data_path] \
 --train_line_path [training_wireframes_path] \
 --mask_path ['irregular_mask_list.txt', 'coco_mask_list.txt'] \
 --train_epoch 12 --validation_path [validation_data_path] \
 --val_line_path [validation_wireframes_path] \
 --valid_mask_path [validation_mask] --nodes 1 --gpus 1 --GPU_ids '0' --AMP

python TSR_train.py --name places2_continous_edgeline --data_path [training_data_path] \
 --train_line_path [training_wireframes_path] \
 --mask_path ['irregular_mask_list.txt', 'coco_mask_list.txt'] \
 --train_epoch 15 --validation_path [validation_data_path] \
 --val_line_path [validation_wireframes_path] \
 --valid_mask_path [validation_mask] --nodes 1 --gpus 1 --GPU_ids '0' --AMP --MaP

Train SSU

We recommend to use the pretrained SSU. You can also train your SSU refered to https://github.com/ewrfcas/StructureUpsampling.

Training LaMa First

python FTR_train.py --nodes 1 --gpus 1 --GPU_ids '0' --path ./ckpt/lama_places2 \
--config_file ./config_list/config_LAMA.yml --lama

Training FTR

256:

python FTR_train.py --nodes 1 --gpus 2 --GPU_ids '0,1' --path ./ckpt/places2 \
--config_file ./config_list/config_ZITS_places2.yml --DDP

256~512:

python FTR_train.py --nodes 1 --gpus 2 --GPU_ids '0,1' --path ./ckpt/places2_HR \
--config_file ./config_list/config_ZITS_HR_places2.yml --DDP

More 1K Results

Acknowledgments

• This repo is built upon MST, ICT and LaMa.

Cite

If you found our program helpful, please consider citing:

@inproceedings{dong2022incremental,
      title={Incremental Transformer Structure Enhanced Image Inpainting with Masking Positional Encoding}, 
      author={Qiaole Dong and Chenjie Cao and Yanwei Fu},
      booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
      year={2022}
}