Pytorch transforms based on numpy arrays. I decided to re-write some of the standard pytorch transforms using only numpy operations that allow for High Dynamic Range image manipulation. The file exr_data.py include some methods for loading HDR images in exr format into numpy arrays and writing numpy arrays into exr files.

A list of implemented transforms based on numpy arrays are:

• Bilateral Filter
• Median Filter
• Image Rotation (either randomly sample an angle between two bounds or with a fixed angle)
• Random Horizontal Flip
• Random Vertical Flip
• Random Crop
• Center Crop
• Five Crops (4 courners + center)
• Normalize 0-1 (Normalize the image between 0-1)
• Random Erasing (Random Erasing Data Augmentation by Zhong et al.)
• To Tensor
• rgb2xyz (Change in the color space)
• xyz2rgb (The opposite change in color space)
• Lambda (Based on torchvision.transforms.Lambda)
• Compose (Based on torchvision.transforms.Compose)
• Normalize (Based on torchvision.transforms.Normalize)

• numpy
• torch http://pytorch.org/
• torchvision http://pytorch.org/
• scipy
• (Only if you want to load/ save HDR images) OpenEXR http://www.excamera.com/sphinx/articles-openexr.html

Create a dataset that loads hdr images in .exr format:

import exr_data

trf = np_transforms.Compose([
    np_transforms.Scale(size=(256, 256)),
    np_transforms.RandomCrop(size=(224, 224)),
    np_transforms.RandomVerticalFlip(prob=0.5),
    np_transforms.RandomHorizontalFlip(prob=0.5),
    np_transforms.RotateImage(angles=(-15, 15)),
    np_transforms.ToTensor(),
])

data_train = exr_data.exrData(root=os.path.join(ROOT_DIR, 'train'),
                                  loader=exr_data.exr_loader,
                                  transform=trf)

                                  

Write a numpy array into a .exr file:

import exr_data

# let's assume we have a numpy array called 'pic' with the image stored in the form [HxWxC]

(Rs, Gs, Bs) = [pic[:, :, channel].tostring() for channel in range(pic.shape[-1])]
exr_data.exr_writer(out_path, size=im_size, pixel_values={'R': Rs, 'G': Gs, 'B': Bs})