Key Concepts

Annotation:

An annotation is defined as the combination of a label attached to a bounding boxed positioned on an image. Per default, any incoming image has one bounding box on the entire image with the label that was attached at time of upload. You can modify the position of the default bounding box or add additional annotations to the image using the Annotation Queue.

Collection Modes:

There are 4 possible collection modes of sources: Train Data for Recognition, Test Data for Recognition, Train Data for Background, Test Data for Background. The mode of a source is immutable.

a) Train Data for Recognition are sources that you plan to use for training your model on the labels that you want the SDK to recognize.

b) Test Data for Recognition are sources that you plan to use for testing your model on the labels that you want the SDK to recognize.

a) Train Data for Background are sources that you plan to use for training your model on background - images that do not contain any of the labels that you want the SDK to recognize.

b) Test Data for Background are sources that you plan to use for testing your model on background.

Dataset:

Datasets are snapshots of the current data. They do not contain the actual images, but point to the images by referencing them. Deleting images on the platform will also delete them for any datasets that are pointing to them. This will not result in an error when retraining on these datasets, but the images will not be included anymore in the training, since they have been deleted.

Sources:

A source is a set of data defined by a unique set of data collection parameters and settings that justify to separate it from other sets of data.