cvtorch: Utilities for pytroch based on opencv

This repository is intended to offer some common augmentataion functions in computer vision task base on opencv. The functions' prototype comes from FAIR's maskrcnn-benchmark. I tried my best to implement these functions strictly follow the details of torchvision and pillow. Any discussions are welcomed.

Modules

• cvtorch.cvBox.BoxList
  This is a handy structure in object detection implemented by FAIR. original code

• cvtorch.cvTransforms.Compose
  Compose a series of transforms to an opencv image (numpy ndarray in BGR order) and cvBox target

• cvtorch.cvTransforms.Resize
  Resize an opencv image and target. target is optional.

• cvtorch.cvTransforms.RandomHorizontalFlip
  random horizontally filp an opencv image and target. target is optional.

• cvtorch.cvTransforms.RandomVerticalFlip
  random vertically filp an opencv image and target. target is optional.

• cvtorch.cvTransforms.ColorJitter
  randomly jit brightness, contrast, saturation, hue of an opencv image and targe. target is optional.

• cvtorch.cvFunctional.ColorJitter
  randomly jit brightness, contrast, saturation, hue of an opencv image.

• cvtorch.cvTransforms.ToTensor
  convert an opencv image to tensor. (follow torchvision with transpose).

• cvtorch.cvTransforms.NormalizeAsNumpy
  Normalize an opencv image.

• cvtorch.cvTransforms.NormalizeAsTorch
  Normalize a tensor image.

Eg

from cvtorch import cvTransforms as T

normalize_transform = T.NormalizeAsTorch(mean, std)
color_jitter = T.ColorJitter(brightness, contrast, saturation, hue)
transforms = T.Compose(
	[
		color_jitter,
		T.Resize(min_size, max_size),
		T.RandomHorizontalFlip(0.5),
		T.RandomVerticalFlip(0.5),
		T.ToTensor(),
		normalize_transform,
	]
)

Functions

• cvtorch.cvFunctional.adjust_brightness

• cvtorch.cvFunctional.adjust_contrast

• cvtorch.cvFunctional.adjust_hue

• cvtorch.cvFunctional.adjust_saturation

installation

pip install cvtorch