Key concepts

Below follows a summary of different key concepts that are used commonly in the Insights product.

The taxonomy specifies what objects should be annotated. For example, it often defines what kind of shapes should be used to annotate. Another key part of the taxonomy is the set of classes (or properties in general) that shapes can be annotated with.

A taxonomy is usually accompanied by a guideline that helps to decide in which cases to annotate or not, what kind of property value to choose in doubt - and how to annotate across a wide range of special cases.

In some cases, Kognic might create two taxonomies to determine "how" annotations are created - for example, by splitting the annotation task into two smaller taxonomies instead of one big. This usually should not change the structure of the output you receive, but it does create two annotations in our platform, whereas eventually, you export only one. For this reason, we refer to the taxonomy also as the Task Definition.

A group of inputs. An input is a set of input data that you want to be annotated. It usually consists at least of one image. It can also contain images from several cameras, a LiDAR point cloud, a video, or a sequence of images.

See also Batch in Kognic API Docs

An annotation is usually a set of shapes, drawn according to the defined taxonomy, on top of one unit of input data uploaded by you. 

Currently, only annotations with the status “Delivery Ready” can be viewed in Insights.

One unit of input data might result in multiple parallel Delivery Ready annotations, if it was used in multiple taxonomies respectively task definitions.

One or multiple shapes with the same unique id inside one annotation. It can be represented in multiple sources (cameras, LiDARs) and in multiple frames of one sequence. An object in this context is mostly a 'thing' (car, person, traffic sign, ..) but can also be 'stuff' (road, debris, ..).  

If a specific car is visible in two sequence annotations, it's one object in the first sequence and another object in the second sequence - as its shapes have different id:s.

One instance of an object is its appearance in one frame and one sensor. 

Examples:

A scene is the information of a physical environment, captured via sensors. This can span a single time instance or a time period as a sequence of information.

At Kognic, this means it's a collection of multiple images, point clouds or other sensor information for one or multiple time instances. A scene is part of an Input.