Package implementing some common function used when performing data augmentation to train deep optical flow networks in PyTorch. The transformations are implemented directly in PyTorch, and they can operate over batches of images.
Download and put flow_transforms.py somewhere it can be accessed from your
main project (e.g. in the same folder). Then just import it and use its functions.
A simple example is show below. A more complete example is shown in the accompanying
Jupyter notebook demo.
import flow_transforms
image1 = function_to_read_image()
image2 = function_to_read_image()
flow = function_to_read_flow()
transform = flow_transforms.Compose([
flow_transforms.ToTensor(), # Must come before other transforms
flow_transforms.RandomHorizontalFlip(),
flow_transforms.RandomAdditiveColor(stdev=1.0)
])
images_tensor, flow_tensor = transform([image1, image2], [flow])The example above is generic and simple just for explanation purposes. But usually you should use these transforms as part of the PyTorch data loading pipeline.
This code is licensed under the MIT License.
Many of the transforms were adapted from https://github.com/ClementPinard/FlowNetPytorch.
The script flowviz.py was taken from https://github.com/georgegach/flowiz.