AnnotateIt 6.16.12
Curated auto-annotation models arrive — ready-to-use detectors D-FINE, RT-DETR, RT-DETRv2, DEIM and RF-DETR, plus EdgeCrafter ECDet and the EdgeCrafter ECSeg instance segmenter — downloaded from the Models page and run locally on your CPU. Use one to pre-label a whole dataset, a selection, a filter or a search in batch, and review the drafts in a Pending AI Review queue before they count. SAM 3 Tracker joins the one-click mask engines, with three new sample projects to start from.
Added
- Camera Capture can now record silent video clips as well as photos. The recorder chooses a supported MP4 or WebM type on the device, stages the completed clip locally for review, and adds it to the dataset only after Accept.
- Auto-annotation models: a set of curated, ready-to-use object detectors — D-FINE, RT-DETR, RT-DETRv2, DEIM and RF-DETR — plus EdgeCrafter ECDet, and EdgeCrafter ECSeg for instance segmentation. Download one from the Models page and set it up inside a project (map labels, test on one of your images, save). Every model runs locally on your CPU and is trained on the 80 COCO classes.
- Batch auto-annotation from the dataset toolbar: pre-label a whole dataset, a selection, a filter or your search results with a set-up model, a text prompt or a visual prompt. Results wait in a Pending AI Review queue — kept out of exports, dataset versions and statistics — until you accept them, with a Review button that stays in the toolbar until the queue is empty.
- SAM 3 Tracker added to the one-click mask engines, alongside MobileSAM and the SAM 2.1 variants (WebGPU required; you accept Meta’s SAM License before the download starts).
- Three new one-click sample projects: Traffic Jam (detection), Wildlife Survey (detection) and Dinosaur Hunting (instance segmentation).
Improved
- The project’s prediction-models tab is redesigned as compact cards — name, a one-line characteristic, size and an info popover — with a single, clearly-marked Recommended model.
- The Models page now presents the curated detectors and segmenters as branded families with Recommended and higher-capacity variants, instead of listing them as generic custom-ONNX imports.
- Setting up a curated model verifies the downloaded file against its published specification on your device before saving, so a wrong or corrupt file is caught during setup rather than mislabelling predictions.
Fixed
- Deleting a media item, or clearing all annotations on one, no longer fails with a ‘reading map’ error.
- The offline prediction source no longer shows a phantom selected model; inference waits for a real, fully-loaded source before it will run.
- Switching the active custom model no longer leaks the previous model’s inference session.
These notes track the web app at app.annotateit.ai, which is where a release goes live first. The Windows and macOS builds are the same product and pick up this work once each store has reviewed it, on their own schedule; the iPhone & iPad builds run a reduced feature set, so not everything here reaches them at all. Every store shows the version it is currently offering, and that can be well behind the web version. What differs by platform.