Instance segmentation
Click an object and Segment Anything turns it into an editable polygon by default. Refine it with more clicks, or switch the output to Mask when you want the model’s exact pixels and brush editing — all on your own hardware.
Related video tutorial
ECSeg-M instance masks in seconds, on-device
Click once and get a mask. Add positive and negative clicks until the outline is right.
Choose Mask output to store the model’s exact pixels rather than a traced outline. COCO and Datumaro receive it as run-length encoding, so the mask survives the round trip exactly; Polygon remains the editable default.
Paint a mask by hand, or clean up the edges of a generated one, with an adjustable brush.
Draw a rough box around an object and let the model fill in the exact shape.
For the cases a model gets wrong, draw and edit polygons point by point — and rotate a finished one with a handle.
Trace something that is a line rather than a region — a lane, a wire, a crack. Datumaro keeps the raw line; COCO and YOLO receive it as a polygon ribbon of the width you choose, and Pascal VOC, being detection-only, gets that ribbon’s bounding box.
Not every mask needs a click. EdgeCrafter ECSeg — the curated instance-segmentation model — downloads from the Models page, maps its 80 COCO classes to your labels and proposes instance masks across a whole image. It is verified on your device before it saves and runs on your CPU; its proposals arrive as drafts you accept or reject.
MobileSAM is built in and ready immediately. Optional SAM 2.1 Tiny, Small and Large and SAM 3 Tracker appear when the platform supports their runtime; every available engine still runs locally.
AnnotateIt now hosts EdgeCrafter ECSeg, a curated instance-segmentation model you download from the Models page and set up in your project: map its 80 COCO classes to your labels, test, save. It runs on your CPU, and a batch run can pre-label a whole dataset — every mask held in a review queue until you accept it.
Use images, native video or Camera media and keep each instance as an object-mask track across video keyframes. COCO and Datumaro both carry a mask natively, as run-length encoding, so the pixels survive the round trip exactly. YOLO has no raster type, so a mask is traced into polygon rows on the way out — which loses holes and thin bridges; Pascal VOC is detection-only and receives the bounding box. Both are documented behaviour, and the export dialog says which one you are about to get before you commit.
Five engines answer the same click, and they are not five sizes of the same thing. Each row below is the trade it actually makes; switch between them any time on the Models page, and the whole toolbar follows.
The engine that works the moment the app opens, on every platform — including iPhone and iPad, where it is the only one offered. Upstream it is SAM with the heavy ViT-H image encoder swapped for a distilled TinyViT; the prompt encoder and mask decoder are SAM’s own. That makes it by far the fastest to encode an image and the quickest click-to-mask — the trade is coarser edges on cluttered scenes and fine structures.
The recommended upgrade. Meta’s SAM 2.1 architecture at its smallest, noticeably more accurate than MobileSAM on object boundaries, and it still runs on the CPU path — so it works in the packaged Windows and macOS apps and in browsers without WebGPU. With a WebGPU browser the image encode drops to a fraction of a second.
Sharper masks than Tiny, from a quantised build that is actually the smaller download of the two. The cost is a slower click. Pick it when boundary quality on stills matters more than response time — and when you want that without the WebGPU requirement the two heavyweight engines carry.
The highest-quality SAM 2.1, and the first row where the runtime narrows: it needs the app’s WebGPU capability, so it is offered in Windows, macOS and compatible desktop browsers with a working WebGPU adapter, and is unavailable on mobile. Encoding an image takes seconds, not milliseconds — worth it on genuinely hard frames, wrong for rapid-fire labelling.
A community ONNX build of Meta’s SAM 3 (the Tracker/PVS variant, used here on single images). The most accurate engine in the row, and — counter-intuitively — it encodes several times faster than SAM 2.1 Large in our benchmark. Also WebGPU-only, and the one engine with a licence step: you review and accept Meta’s SAM License before the download starts.
Two mechanics explain most of what you feel. The image encoder runs once per image and is where the seconds go; every click after that reuses the embedding, so refining a mask stays in the tens of milliseconds on any engine. And the mask decoder deliberately stays on the CPU path even when WebGPU is available — the WebGPU decoder route produces wrong masks in the current ONNX Runtime build, so only the encoder is accelerated.
On an Apple M4 Max with WebGPU, one image encodes in about 72 ms on MobileSAM, 150 ms on SAM 2.1 Tiny, 1.9 s on Small, 3.3 s on SAM 3 Tracker and 10.3 s on Large; clicks land in 12–42 ms on all five. Latency is measured; which engine leaves you the least correction work on your images is not — test the free ones on your own data before paying the download for a heavyweight.
MobileSAM is bundled and works everywhere. SAM 2.1 Tiny and Small are optional downloads on supported web and desktop builds with a CPU fallback. SAM 2.1 Large and SAM 3 Tracker require the app’s WebGPU capability, so they are offered in Windows, macOS and compatible desktop browsers with a working WebGPU adapter. They remain unavailable on mobile. Every available engine runs on your device, and Models shows its licence, size, speed and accuracy first.
No. The image is encoded and segmented locally, in your browser or in the native app. Nothing is uploaded.
Yes. Refine it with more clicks, paint over it with the brush, erase part of it, or fall back to manual polygon editing at any point.
Yes — EdgeCrafter ECSeg, the curated instance-segmentation model. Download it from the Models page, map its 80 COCO classes to your labels, and it proposes instance masks for the whole image rather than one click at a time. It is verified on your device before it saves and runs on your CPU; its proposals arrive as drafts you accept or reject, so nothing is written behind your back.
Yes. Start Auto-annotate over the whole dataset, a selection, the current filter or semantic-search results, using Text Prompt, Find by example or a set-up segmentation model. Predictions wait in Pending AI Review and images with no match are reported separately; nothing is written as final until you accept it.
Yes — EdgeCrafter ECSeg, a curated instance-segmentation model trained on the 80 COCO classes. Download it from the Models page, map its classes to your labels and test it on one of your images before saving. It runs on your CPU, and a batch run pre-labels a whole dataset into a review queue that stays out of your exports and versions until you accept it.
Create your first local project in the browser. Optional ChatGPT or Claude annotation needs your own AI connection.
Questions before you start?Contact support →