Label the objects in your photos

Why class names matter

A model learns only what the labels tell it. If one photo says “car”, another says “Car” and a third says “cars”, a training script sees three classes. Most labelling mistakes in small datasets are of this kind, not badly drawn outlines. The class bar in CutLift is there to prevent them: you type a class name once, give it a colour, and from then on you pick it rather than retype it.

Labelling quickly with the class bar

The first nine classes have number keys. Select one or more layers and press 1 to 9 to label them, or press a number with nothing selected and every new object you mark takes that class. Select all of this class gathers every object with the same label, so you can check them together, and Colour layers by class paints each object in its class colour, which makes a mislabelled object stand out at a glance.

Rename or recolour a class at any time; the objects that use it follow. Classes can be reordered so the ones you use most sit under the first keys.

Tag objects in images, not just whole photos

This is a data labeling tool for objects: each label belongs to an outline, not to the photo as a whole. Pick each object with a lasso, Quick Selection or the Pen, or label many at once with Select by name (Pro), which selects every match for a word you type.

Labeling for machine learning at object level tells a model where each thing is, not only that it appears somewhere in the picture. Areas you do not label are simply background: they get no annotation in COCO and the value 0 in a label map.

From labels to training files

Labelled photos export as one mask per object and a Simple JSON for free. With Pro they also export as COCO JSON, YOLO segmentation labels or an index label map. Your class list carries over from photo to photo, so a class keeps the same id in every export of a set, and you can load an existing dataset’s classes.txt, data.yaml or COCO file to match its ids. Labelling is done in your browser, and nothing you tag is sent to a server.