ParticleDetection.modelling.datasets

Functions to prepare datasets for the use by the Detectron2 framework, as well as functions to get basic information about a dataset, i.e. size and thing classes.

Author: Adrian Niemann (adrian.niemann@ovgu.de)

Date: 31.10.2022

extract_polygon(annotation: dict)

Extract a polygon and its bounds from annotation data.

This function extracts object segmentations as polygons from different shapes annotated with the VGG Image Annotator (VIA). Currently this function supports rectangular, circlular, elliptical, and polyonal/polyline annotations from VIA.

Parameters:

annotation (dict) – Contents of the "shape_attributes" field of an object’s segmentation data saved from VIA.

Returns:

tuple(list, list) – [0] : list of polygon point coordinates

[1] : bounding box of object as [min_x, min_y, max_x, max_y]

Raises:

ValueError – Is raised, in case an unknown annotation type is encountered, i.e. none of the ones mentioned above.

get_dataset_classes(dataset: DataSet) → Set[int]

Retrieve the number and IDs of thing classes in the dataset.

get_dataset_size(dataset: DataSet) → int

Compute the number of annotated images in a DataSet (excluding augmentation).

load_custom_data(dataset: DataSet) → List[dict]

Loads a (training/testing) dataset into the Detectron2 format.

Parameters:

dataset (DataSet) – Dataset that will be transferred to the Detectron2 dataset format for training a model.

Returns:

List[dict] – Dataset in the Detectron2 format.

Note

For more information see: https://detectron2.readthedocs.io/en/latest/tutorials/datasets.html#standard-dataset-dicts

register_dataset(dataset: DataSet, generation_function: Callable = load_custom_data, classes: List[str] | None = None)

Register a custom dataset to the Detectron2 framework.

Parameters:

• dataset (DataSet)

• generation_function (Callable, optional) –

  Function, that transforms a given DataSet into a Detectron2 readable format.

  By default load_custom_data().

• classes (List[str], optional) –

  Names of the classes present in the loaded dataset.

  By default None, which results in class names like 0, 1, 2, ....