Geometry Primitives

Potato's image surface covers bounding boxes, polygons, polylines, ellipses, 2D cuboids, keypoint sets with skeleton topology, and instance-keyed brush masks. All of them are configured through image_annotation's tools list.

annotation_schemes:
  - annotation_type: image_annotation
    name: objects
    description: "Outline everything you find."
    source_field: image
    tools: [bbox, polygon, polyline, ellipse, cuboid_2d, keypoint_set, brush, eraser, sam]
    labels:
      - {name: vehicle, color: "#FF6B6B", key_value: "1"}
      - {name: person, color: "#4ECDC4", key_value: "2"}

Runnable example: examples/image/geometry-primitives/.

The tools

Keypoint sets carry COCO visibility flags and a skeleton topology, so a pose exports as a pose rather than as a bag of points.

Keyboard conventions changed in 2.8

Defaults now follow V7 and CVAT conventions:

This is a change: b used to be bbox and f used to be freeform. A study already in the field can keep the old bindings:

    keybinding_profile: legacy

Annotators see a one-time notice about the change rather than discovering it by drawing the wrong shape.

Working faster

• Per-class show and hide with h and Shift+H, persisted across the whole project rather than reset per image.
• Copy annotations from the previous image with Ctrl/Cmd+D, or set carry_over to do it automatically. Useful for near-duplicate frames.
• Two-finger pan and pinch zoom for tablets and stylus work.

Fixes worth stating

Several image features had shipped in earlier versions without working, and saying so is more useful than quiet correction:

• Accepting an AI detection called a method that did not exist, so it was a silent no-op.
• The freeform tool drew paths that were never saved.
• Brush strokes leaked between instances, through three separate instance-switch bugs.
• Image annotation was completely inert during the training phase.

All are fixed in 2.8. If you ran an image study on an earlier version, the training phase in particular is worth checking.

Related

• Interactive segmentation
• Text prompting
• Geometry agreement
• Image annotation reference