CCAugmentation

Data preprocessing and augmentation framework that is designed for working with crowd counting datasets. It supports multitude of simple as well as advanced transformations and is based on pipelines which allow a flexible flow of data between loaders, transformations and outputs. Deep learning framework-independent, though works best with PyTorch.

Current capabilities

Each data preprocessing procedure is defined in form of a pipeline that consists of a data loader and a list of operations to sequentially execute on the data. Each of the operations may be of the following types:

• Transformation - Returns transformed data on output, does not have side effects
• Output - Returns unmodified data on output, has side effects that, for example, write data to files
• Operation - Performs any other functions, not qualifying for any of the aforementioned types

Available transformations are:

• Crop
• Scale
• Downscale
• Rotate
• StandardizeSize
• Normalize
• NormalizeDensityMap
• FlipLR
• ToGrayscale
• LambdaTransformation
• Cutout

Available outputs are:

• Demonstrate
• SaveImagesToFiles
• SaveImagesToBinaryFile
• SaveDensityMapsToCSVFiles
• SaveDensityMapsToBinaryFile

Available operations are:

• Duplicate
• Dropout
• RandomArgs
• OptimizeBatch

Available loaders are:

• BasicImageFileLoader
• ImageFileLoader
• BasicGTPointsMatFileLoader
• GTPointsMatFileLoader
• BasicDensityMapCSVFileLoader
• DensityMapCSVFileLoader
• VariableLoader
• ConcatenatingLoader
• CombinedLoader

You can also use builtin integrations for:

• PyTorch
• ShangaiTech dataset
• NWPU dataset

For more information about the specific topics, please refer to the related comments in the code.

How to use

Loading the data from ShanghaiTech dataset and taking crops with 1/4 size:

import CCAugmentation as cca
import CCAugmentation as ccat


train_data_pipeline = cca.Pipeline(
    cca.integrations.loading.SHHLoader("/data/ShanghaiTech/", "train", "B"),
    [
        ccat.Crop(None, None, 1/4, 1/4)
    ]
)

train_img, train_dm = train_data_pipeline.execute_collect()
# you can also use execute_generate() to create a generator

print(len(train_img), len(train_dm))

To see more examples of usage, please see our experiment environment repository.

You can also preview the documentation for this project in a browser, using pdoc3 docs generation. Here's how you can easily do this:

$ pip install pdoc3
$ pdoc --html --output-dir docs CCAugmentation