Label your first images

Create a project

1. Choose File → New Project… (⌘ N).
2. Name it Sample Objects and leave Object detection selected. Segmentation is unavailable.
3. Choose a parent save folder outside your source image folder, then choose the image folder.
4. Enter classes such as bottle and cup, one per line. Choose Create Project and wait for copying to finish.

Folders and subfolders

You can also start with Create New Project… in the project browser. Enable Include subfolders for nested images; it is off by default.

Annotension creates a new project folder and refuses an existing folder with the same name. It contains project.annotension and images/. Your source images stay unchanged.

Draw around each object

For a larger batch, group similar images before labeling to work through related scenes together.

1. Click an image in the gallery. Choose its object class in the sidebar.
2. Choose Edit → Create box tool (⌘ B), then drag across the object.
3. Select a box to move it. Drag a handle to resize it, or choose another class to change its label.

Fit boxes closely around visible objects and use the same labeling rules throughout the dataset. Delete removes the selected box; Edit → Undo reverses a mistake.

Navigation and keyboard shortcuts

Fit image

⌘ 0

Zoom

Pinch or Command-scroll

Pan

Space-drag or scroll

Deselect

Escape

Previous / next

⌘ [ / ⌘ ]

Gallery / editor

⌘ G / ⌘ D

Check the whole image

Inspect every box and class. Look for missed objects. Then check Image reviewed (including empty).

An image with no target objects still needs inspection and review. This makes it a verified negative. Leaving it empty does not approve it.

Editing boxes clears review. Inspect and approve the image again. If you use prediction, resolve all pending suggestions before image approval.

Autosave and reopening

Changes autosave. Resolve any save error before switching projects or quitting. Reopen through File → Open Project… or the project browser.

Export reviewed data

1. Choose File → Export… (⌘ E).
2. For a small first test, choose YOLO labels only.
3. Choose Choose Destination…, then a parent output folder. Annotension creates a fresh export subfolder.
4. Wait for Export Complete and choose Show in Finder.

Only reviewed images are included. A reviewed negative has an empty label file.

What “labels only” includes

The export includes orientation-corrected PNGs, matching normalized text labels, class names, filename mapping, and an export-time project snapshot. There is no train/validation split.

Training-ready YOLO and COCO exports

Choose YOLO training dataset or RT-DETRv2 / COCO detection for a training split. These require at least two distinct reviewed source files and reviewed object annotations.

Choose a validation share and seed for a new manual project. Supplied splits from imported datasets are preserved. Byte-identical copies stay together, so actual counts can differ from the requested share. Inspect split coverage before training.

Export folder
├── images/         PNG images
├── labels/         YOLO labels
├── classes.txt
├── filenames.json
└── project-snapshot.json