[PBVS @ CVPR 2024 Thermal Images Super-resolution challenge- Track 1]

Our method achieved 4th place in PSNR and 3rd place in SSIM in the PBVS 2024 Thermal Image Super-Resolution Challenge(TISR)-Track1.

Environment

• Python == 3.8.0
• PyTorch == 1.11.0
• BasicSR == 1.3.4.9
• eninops

Installation

• Install Pytorch, BasicSR and Einops first. Then,

  python setup.py develop

  Data Preparation

  DF2K for Pre-training

• Preparation of DF2K dataset can refer to this page. save it in ./datasets

Extract sub-images

python df2k_extract_subimages.py

Create a meta-info file

python df2k_generate_meta_info.py

Thermal Image for Fine-tuning

• Preparation of Thermal Image can refer to this page.

Extract sub-images

python thermal_extract_subimages.py

Create a meta-info file

python thermal_generate_meta_info.py

Quick[test]

• Refer to ./options/test
• The pretrained models are available at Google Drive. Save it in ./experiments/pretrained/. (best performance at 14,000 iterations)

Create SR images

python hat/test.py -opt options/test/HAT_SRx8_quick.yml

The testing results will be saved in the ./results folder.

Pre-training

• Refer to ./options/train Pretraining command

  CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 --master_port=1234 hat/train.py -opt options/train/train_HAT_thermalSRx8_pre.yml --launcher pytorch

  The training logs and weights will be saved in the ./experiments folder.

Fine-tuning

• Refer to ./options/train
• The model pre-trained on the DF2K dataset is available at Google Drive. Save it in ./experiments/pretrained/. (We conducted training for 191,000 iterations.) Fine-tuning command

  CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 --master_port=1234 hat/train.py -opt options/train/train_HAT_thermalSRx8_48_cutblur_fineturn.yml --launcher pytorch

  The training logs and weights will be saved in the ./experiments folder.

Acknowledgment

Our codes borrowed from chxy95 and nmhkahn. Thanks for their work.