AI-assisted annotation for crowd objects, without the duplicates

Cars in a packed parking lot, people in a stadium, colonies on a plate. One image, many instances of the same thing. Call these crowd objects: the hard case for AI-assisted annotation.

A single AI-Box click can return three predictions on the same car: the car itself, the cluster around it, and a hubcap on top. A second AI-Box prompt to pick up the rest of the lot then duplicates every car you already labeled. SAM is doing its job; nothing strips the redundancy before shapes land on the canvas.

LabelMe v6.3 adds two suppression passes that do.

How it works

Nested masks collapse to one shape. A single click can return multiple SAM masks at different granularities: the whole cluster, one instance, a part of that instance. v6.3 keeps the best per label and drops the rest, comparing both mask overlap and how much one mask is contained in another. Plain bbox NMS cannot deduplicate nested masks that share a bounding box; this one can.

Re-clicks over annotated regions don't duplicate. When a prediction overlaps a shape already on the canvas, it's dropped before becoming a shape. The check runs against rectangle, polygon, oriented_rectangle, circle, and mask outputs alike.

Together, a noisy SAM pass turns into a clean, additive one. Only genuinely new objects reach the label list.

When to reach for it

• Aerial parking lots, stadiums, crowds. The hardest case: every visible instance touches at least two others, and SAM happily returns predictions at every scale.
• Retail shelves and dense product layouts. Identical items, repeated and overlapping; the second AI pass no longer fights the first.
• Microscopy. Cells, colonies, particles. Many small instances in one frame, where re-clicking is the natural workflow.
• Any image you'll annotate in multiple AI passes. Different classes, different parameters, different days. The new behavior means later passes are additive instead of duplicative.

Where to find it

Update to LabelMe v6.3 from the download page. The suppression is automatic for AI-Box, AI-Points, and AI-Text, with no settings to toggle.

See the v6.3 release notes for the full changelog. The two suppression passes landed in #2087 (mask-aware NMS) and #2088 (overlap-with-existing suppression).

If anything behaves oddly, open a GitHub issue or ask on Discord.