Dataset representation

Each dataset is represented by a pandas dataframe. Each row corresponds to one entry, which depicts one individual animal. Usually one entry corresponds to one image and one animal. However, sometimes, there are multiple animals in an image and then one image may generate multiple entries differentiated by additional information such as bounding box. Columns are descriptions of the entry.

Required columns

The following three columns must be part of all dataframes.

Column	Type	Description
id	int or str	Unique id of the entry.
identity	int or str	Identity (or label) of the depicted individual animal.
path	str	Relative path to the image.

There is a special value for identity which describes an unknown individual. Its default value for unknown animals is

from wildlife_datasets import datasets

datasets.DatasetFactory.unknown_name

unknown

When a dataset contains unknown inidividuals, there identities should be changed to the default value.

Optional columns

The following columns may be present in the dataframe. Besides these columns, it is possible to define additional columns.

Column	Type	Description
bbox	List[float]	Bounding box in the form [x, y, w, h]. Therefore, the topleft corner has coordinates [x, y], while the bottomright corner has coordinates [x+w, y+h].
date	special	Timestamp of the photo. The preferred format is %Y-%m-%d %H:%M:%S from the datetime package but it is sufficient to be amenable to pd.to_datetime(x).
keypoints	List[float]	Keypoints coordinates in the image such as eyes or joints.
position	str	Position from each the photo was taken. The usual values are left and right.
segmentation	List[float] or special	Segmentation mask in the form [x1, y1, x2, y2, ...]. Additional format are possible such as file path to a mask image, or pytorch RLE.
species	str or List[str]	The depicted species for datasets with multiple species.
video	int	The index of a video.

Metadata

Besides the dataframe, each dataset also contains some metadata. The metadata are saved in a separate csv file, which currently contains the following information. All entries are optional.

Column	Description
name	Name of the dataset.
licenses	License file for the dataset.
licenses_url	URL for the license file.
url	URL for the dataset.
cite	Citation in Google Scholar type of the paper.
animals	List of all animal species in the dataset.
animals_simple	List of all animal species in the dataset.
real_animals	Determines whether the dataset contains real animals.
year	Publication year of the dataset.
reported_n_total	The reported number of total animals.
reported_n_identified	The reported number of identified animals.
reported_n_photos	The reported number of photos.
eported_n_individuals	The reported number of individuals.
wild	Determines whether the photos were taken in the wild.
clear_photos	Determines whether the photos are clear.
pose	Determines whether the photos have one orientation (single), two orientation such as left and right flanks (double) or more (multiple).
unique_pattern	Determines whether the animals have unique features (fur patern, fin shape) for recognition.
from_video	Determines whether the dataset was created from photos or videos.
cropped	Determines whether the photos are cropped.
span	The span of the dataset (the time difference between the last and first photos).
size	Size of the zipped datasets (in MB).