/ Images

OpenCV implementation of Torchvision's image augmentations

OpenCV implementation of Torchvision's image augmentations

opencv_transforms

This repository is intended as a faster drop-in replacement for Pytorch's Torchvision augmentations. This repo uses OpenCV for fast image augmentation for PyTorch computer vision pipelines.

Requirements

  • A working installation of OpenCV. Tested with OpenCV version 3.4.1
  • Tested on Windows 10. There is evidence that OpenCV doesn't work well with multithreading on Linux / MacOS, for example num_workers >0 in a pytorch DataLoader. I haven't tried this on those systems.

Installation

  • git clone https://github.com/jbohnslav/opencv_transforms.git
  • Add to your python path

Usage

  • from opencv_transforms import opencv_transforms as transforms
  • From here, almost everything should work exactly as the original transforms.

Example: Image resizing

import numpy as np
image = np.random.randint(low=0, high=255, size=(1024, 2048, 3))
resize = transforms.Resize(size=(256,256))
image = resize(image)

Should be 1.5 to 10 times faster than PIL. See benchmarks

Performance

  • Most transformations are between 1.5X and ~4X faster in OpenCV. Large image resizes are up to 10 times faster in OpenCV.
  • To reproduce the following benchmarks, download the Cityscapes dataset.
  • An example benchmarking file can be found in the notebook bencharming_v2.ipynb I wrapped the Cityscapes default directories with a HDF5 file for even faster reading.

benchmarking_Resize

benchmarking_Random_crop_quarter_size

benchmarking_Color_brightness_only

benchmarking_Color_constrast_and_brightness

benchmarking_Color_contrast_only

benchmarking_Random_horizontal_flip

The changes start to add up when you compose multiple transformations together.

benchmarking_Resize_flip_brightness_contrast_rotate

TODO

  • [x] Initial commit with all currently implemented torchvision transforms
  • [x] Cityscapes benchmarks
  • [ ] Make the resample flag on RandomRotation, RandomAffine actually do something
  • [ ] Speed up augmentation in saturation and hue. Currently, fastest way is to convert to a PIL image, perform same augmentation as Torchvision, then convert back to np.ndarray

GitHub