Label Studio Vision Labels

A configurable labeling interface for images and video, so boxes, masks and relations export in standard formats.

Label Studio wraps a 10 by 20 box into an export row

Why Does This Exist?

Labeled images decide model quality, but labeling tools either lock formats or need heavy setup. Label Studio sits in the middle: a config file declares the interface, annotators draw boxes, polygons and relations in a browser, and exports land in COCO and other standard forms. It feeds detection and instance segmentation with reviewable human labels.

This page covers the label contract, one area fixture, and appropriate use.

Think of It Like This

A print shop order form

Customers tick boxes on one order form: size, paper, copies. The shop prints exactly what the form declares, and the form files beside the job for reprints.

Label configs are the order form. Annotators fill it per image, exports are the print run. The analogy stops at judgment: forms never argue, while ambiguous objects need written guidelines plus reviewer agreement checks.

How It Actually Works

Input is unlabeled images plus an interface config naming the labels and shapes. Output is annotation rows with coordinates reviewers can accept or reject. Purpose is fast, reviewable human labeling with machine pre-labels optional. Appropriate use is small-to-mid datasets needing custom ontologies and agreement tracking. Write the ontology before labeling starts: renaming a class halfway splits the dataset into two dialects.

Area fixture

Take box $[5, 5, 15, 25]$ in $x_1, y_1, x_2, y_2$ form. Width is $15 - 5 = 10$ , height is $25 - 5 = 20$ , area is $10 \times 20 = 200$ pixels. The export row carries those four numbers plus the label, which is the unit the diagram wraps.

Code

Area in plain Python, the check before export:

1def box_area(x1: float, y1: float, x2: float, y2: float) -> float:2    return max(x2 - x1, 0.0) * max(y2 - y1, 0.0)  # xyxy to pixels3
4
5print(box_area(5.0, 5.0, 15.0, 25.0))  # -> 200.0

200 pixels labeled and exported as one row.

Watch Out For

Changing the ontology mid-project

Adding or renaming a class halfway leaves half the data in the old dialect. The symptom is a model that confuses the renamed class with background. Fix it by freezing the label config, versioning changes, and re-reviewing affected images.

The Quick Version

Label Studio maps images plus a label config to exportable annotation rows.
Box 5, 5, 15, 25 spans 10 by 20 for exactly 200 pixels.
Exports land in standard forms reviewers can check.
Use it for custom ontologies with agreement tracking.
Freeze the ontology or output splits into dialects.