CDTNet: High-Resolution Image Harmonization via Collaborative Dual Transformations (CVPR 2022)

Unofficial implementation of "High-Resolution Image Harmonization via Collaborative Dual Transformations (CVPR 2022)" in PyTorch.

Example Results

Prerequisites

• Linux
• Python 3
• CPU or NVIDIA GPU + CUDA CuDNN

Datasets

• HAdobe5k

Train

1. Download HRNet-W18-C model(hrnetv2_w18_imagenet_pretrained.pth) in HRNets
2. Put it in the pretrained folder.
3. Run:

CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" python train.py --model iih_base --name iih_base_allidh_test --dataset_root ~/IHD/ --dataset_name HAdobe5k --batch_size 80 --init_port 50000

Test

Train a new model

If you train a new model with the command above, latest_net_G.pth and latest_net_P2P.pth will be generated in the directory checkpoints/iih_base_lt_allihd.

Run:

CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" python test.py --model iih_base --name iih_base_allidh_test --dataset_root ~/IHD/ --dataset_name HAdobe5k --batch_size 80 --init_port 50000

Apply pre-trained model

1. Download pre-trained models (model_G and model_P2P)
2. Change their names to latest_net_G.pth and latest_net_P2P.pth respectively.
3. Put them in the directory checkpoints/iih_base_lt_allihd.
4. Run (use the command above)

Quantitative Result

Image Size	CDTNet (officially reported)	CDTNet (implemented)
256x256	38.24	37.42
1024x1024 (after LUT)	37.65	37.13
1024x1024	38.77	37.30

Acknowledgement

We borrowed some of the data modules and model functions from repo of IntrinsicHarmony, iSSAM, and 3DLUT.