An image annotation tool for training data

Who it suits

CutLift is an image annotation tool for people who label their own data: an ML engineer adding a few hundred examples of a new class, a researcher preparing images for a paper, a student training a first segmentation model, or a company whose images cannot be shared with an outside service. It is not a labelling marketplace or a crowd platform. You open a photo, annotate it in the browser and export the files.

Because every label is a layer, an annotation can be reshaped or given a new class long after it was drawn, which matters when a class definition changes halfway through a project.

What one annotation holds

Each object you mark becomes a layer with a pixel mask, a name and a class. On export those layers turn into training data annotation files. Mask PNGs and a Simple JSON file, with a label, a box and a mask for each object, are free.

With Pro you can also export COCO JSON, where each object carries a polygon or RLE outline, a bounding box, its area and its category; YOLO segmentation labels with a classes.txt; an index label map; and SVG outlines.

A steady rhythm for each image

Pick a class in the class bar first, then mark every object of that class before moving on. Number keys 1 to 9 apply a class to the selected layers, or set it for the next objects you mark. Busy scenes go faster with Segment everything (Pro), which proposes a layer for each object: delete the ones you do not want and set the class on the rest.

Before exporting, check small objects at full zoom. Select tiny pieces finds leftover fragments, and Split pieces turns a selection that covers two separate objects into one layer per piece, so each gets its own annotation.

Annotation that stays on your machine

Online image annotation usually means uploading your images to someone else’s server. Here the photos are read and labelled in your browser, which matters for client, industrial or unpublished data. Save a project file to keep every layer and class, and come back to an image when you need to correct it.