FEATURE

AI annotation outputs circles and rectangles, not just polygons

SAM gives you a mask. Your downstream model wants a bounding box, or a rotated box that hugs an off-angle object, or a circle around a cell. Until now, the path was: click to get a polygon, export, write a script that fits the shape you actually need, hope the conversion doesn't lose anything important.

LabelMe v6.2 makes the output shape a property of the AI tool itself. Pick the shape you want; the model fits it for you.

How it works

AI-Points and AI-Box both have an output-shape dropdown next to the model selector. The previous options were polygon and mask. v6.2 adds three more.

rectangle — the axis-aligned bounding box of the SAM mask. Same shape every detector expects, but produced by clicking a point or dragging a box instead of drawing the rectangle by hand.

oriented_rectangle — a minimum-area rotated bounding box fitted to the mask. The box rotates to match the object, so it stops wasting area on diagonals.

circle — a circle fitted to the mask. Useful for round objects where a box overstates the footprint.

Pick once at the top of a session; every AI click produces that shape until you change it.

When to pick which shape

  • Axis-aligned rectangle — anything you're feeding into a standard object detector. YOLO, DETR, Faster R-CNN: they all want axis-aligned boxes.
  • oriented_rectangle — parking lots from above, ships, packaged goods on a conveyor, text in scanned documents. Anything photographed off-angle where an axis-aligned box would include half the neighbor.
  • circle — storage tanks, cells, wells, coins, traffic signs viewed head-on. A single radius is a tighter description than four box edges.
  • polygon — when you want the silhouette and intend to train segmentation, or to keep the option to derive other shapes later.
  • mask — when downstream code expects a raster, not vertices.

Oriented boxes for cars, end to end

Aerial parking-lot footage is the canonical case where axis-aligned boxes fall apart: cars at every heading, neighbors a few pixels away, axis-aligned boxes spilling into the next stall. Here's AI-Box with SAM3 producing oriented rectangles across a parking lot, one click per car.

Where to find it

Update to LabelMe v6.2 from the downloads page. The new outputs appear in the AI-Assisted Annotation dropdown for both AI-Points and AI-Box.

See the v6.2 release notes for the full changelog, including the new oriented rectangle shape that's also available as a hand-drawn tool.

If something behaves oddly with a particular output shape, open a GitHub issue or ask on Discord.

LabelMe is an offline-first annotation tool with built-in AI.

Try LabelMe Free