Project repo for the paper SILT: Self-supervised Lighting Transfer Using Implicit Image Decomposition (BMVC 2021)

The SILT model is written in PyTorch and during the experiments PyTorch 1.8.1 was used with CUDA 11.1 (docker image: nvidia/cuda:11.1.1-cudnn8-runtime-ubuntu18.04). The model was trained on a single GeForce RTX 3090 GPU.

We release the train and test code as well as the pretrained models. The checkpoints of SILT models trained on both datasets discussed in the paper can be downloaded from here.

To test the SILT model, create a checkpoints directory within SILT. Then, within checkpoints create a folder with the name of your experiment, e.g. checkpoints/EXP_NAME. Place the downloaded checkpoint within your experiment folder and rename it to latest_net_G.pth.

To access the checkpoint folder, you have to update the --checkpoints_dir argument in tools/args.py to your path.

Then, to test the model, use the below command:

python test.py --name 'EXP_NAME'

The SILT project defaults to the Multi-Illumination dataset but you can create and place your own datasets in tools/datasets/NEW_DATASET.py and then add them to the dataloader tools/dataloader.py.

To train the model, use the below command:

python train.py --name 'EXP_NAME'

You can also add new datasets as suggested in the Testing section. Please set your checkpointing directory --checkpoints_dir in tools/args.py. To train in a self-supervised manner (with misaligned data), use the --misaligned flag. If you wish to use tensorboard logging, use the --tb_log flag and set your logging directory using the --tb_dir argument.

If you use or write about SILT, please use the below citation:

@inproceedings{kubiak_2021_silt,
  title={SILT: Self-supervised Lighting Transfer Using Implicit Image Decomposition},
  author={Nikolina Kubiak and Armin Mustafa and Graeme Phillipson and Stephen Jolly and Simon Hadfield},
  booktitle={Proceedings of the British Machine Vision Conference (BMVC)},
  year={2021}
}