Extended CycleGAN for RGB-to-polarimetric image transfer. This anonymous repository is associated to the ACCV submission "Generating Polarimetric-encoded Images using Constrained Cycle-Consistent Generative Adversarial Networks" and is here mainly for reproducibility purposes.
The code for the polarimetric RetinaNet if from rachelblin's github, which is itself adapted from fizyr's github.
Anonymous repository: [https://anonymous.4open.science/r/4a83820e-9c65-417c-af3a-ab2979d6e2e8/]
- Supports up to 2 GPUs
- Outputs a TensorBoard log file and the model checkpoints in the 'runs' directory
- The model checkpoints are at the Keras hdf5 saved model format. To load a model, use keras.models.load_model(path_to_model)
See polarcycle_config.py and vanilla_cyclegan_config.py
python -m deeplauncher --config_path config_file --datasets-paths rgb_path polar_pathTakes around two days to train on 2 NVidia GTX1080Ti for the 2485 images datasets.
python -m deeplauncher --config_path config_file --epoch epoch --resume_path --datasets-paths rgb_path polar_pathpython scripts/generate_samples.py checkpoint_path files_path output_pathNeeds keras_retinanet to restore the model
python scripts/evaluate_fid.py fid_model real_imgs fake_imgs- cyclegan_genRGBtoPolar_399.hdf5: Vanilla CycleGAN RGB to polarimetric generator
- polarcycle_genRGBtoPolar_399.hdf5: Our extended CycleGAN RGB to polarimetric generator
- python3
- numpy
- tensorflow >= 1.10
- python-opencv
- progressbar2
- deeplauncher
- keras_retinanet (For computing the FID only)