polar_augment

Description

This repository provides polarimetric image augmentations such as an SO(2) rotation.

Example showing polarimetric azimuth maps after SO(2) rotation

spatial-only rotation	spatial + polarimetric rotation	ground truth

Citation

@ARTICLE{11202388,
  author={Hahne, Christopher and Rodríguez-Núñez, Omar and Gros, Éléa and Lucas, Théotim and Hewer, Ekkehard and Novikova, Tatiana and Maragkou, Theoni and Schucht, Philippe and McKinley, Richard},
  journal={IEEE Transactions on Image Processing}, 
  title={Physically Consistent Image Augmentation for Deep Learning in Mueller Matrix Polarimetry}, 
  year={2025},
  volume={34},
  number={},
  pages={6953-6962},
  keywords={Imaging;Polarimetry;Deep learning;Data augmentation;Training;Optical polarization;Optical imaging;Vectors;Interpolation;Standards;Augmentation;polarimetry;Mueller matrix;tumor;classification},
  doi={10.1109/TIP.2025.3618390}
}

Installation

$ git clone github.com/hahnec/polar_augment
$ cd polar_augment
$ bash install.sh

Usage

The provided transforms expect the image dimensions to be in PyTorch style CxHxW.

import torch

from polar_augment.padding import mirror_rotate

# direct application
from polar_augment.rotation_mm import RandomMuellerRotation
rotate = RandomMuellerRotation(degrees=45, p=float('inf'), pad_rotate=mirror_rotate)
mm_img = torch.randn([128, 128, 4, 4]).flatten(2, 3).permute(2, 0, 1)
mm_img_rotated = rotate(mm_img)
print(mm_img_rotated.shape)

# application for calibration matrices (dataloader-friendly for raw data)
from polar_augment.rotation_raw import RandomPolarRotation
rotate = RandomPolarRotation(degrees=45, p=float('inf'), pad_rotate=mirror_rotate)
mm_img = torch.randn([128, 128, 4*3, 4]).flatten(2, 3).permute(2, 0, 1)
mm_img_rotated = rotate(mm_img)
print(mm_img_rotated.shape)

Alternatively, the transforms can be integrated during dataloading as for example by

from torchvision.transforms import ToTensor
from polar_augment.flip_raw import RandomPolarFlip
from polar_augment.rotation_raw import RandomPolarRotation
from polar_augment.padding import mirror_rotate
from polar_dataset import PolarimetryDataset

# define list of transforms
transforms = [
        ToTensor(), 
        RandomPolarRotation(degrees=180, p=.5, pad_rotate=mirror_rotate), # rotation
        RandomPolarFlip(orientation=0, p=.5),   # horizontal flip
        RandomPolarFlip(orientation=1, p=.5),   # vertical flip
        RandomPolarFlip(orientation=2, p=.5),   # combined horizontal and vertical flip
    ]

# pass transforms to your dataset
PolarimetryDataset(some_file_path, transforms=transforms)

where the augmentation can then be applied to frames and labels within the PolarimetryDataset

class PolarimetryDataset(Dataset):
    def __init__(
            self, 
            path, 
            transforms=[], 
        ):
        self.load_file_paths(path)
        self.transforms = transforms

    def __getitem__(self, i):

        frame = load_file(self.frame_paths[i])
        label = load_file(self.label_paths[i])

        for transform in self.transforms:
            frame, label = transform(frame, label=label)