Annotension

Label. Train.
Predict. Correct.
Repeat.

Label a starting dataset, train a detector on your Mac, and predict on more images. Review and correct the suggestions, grow your dataset, and train again—all locally.

Follow the local loop →Start labeling →
Native macOS app · Local training · Human review
The local annotation and training loop Annotate a starting dataset, train locally, predict on more images, then review and correct the suggestions. Add reviewed data to the next training run and repeat. AnnotateTrain locallyPredict moreReview & correct Reviewed starting labelsOn your MacNew image batchesApprove each imageMore reviewed datafor the next run Human review stays in every cycle
Annotate → Train → Predict → Correct → Repeat

Grow your dataset through review

  1. Annotate →

    Label and review your starting images.

  2. Train locally →

    Train a detector from reviewed labels.

  3. Predict more →

    Add images and generate suggestions.

  4. Correct →

    Fix boxes, classes, and missed objects.

  5. Repeat ↻

    Train again with the expanded reviewed dataset.

Suggestions stay pending until you decide what to keep. Inspect the whole image before approving it, including images where the model finds nothing.

Explore the local loop ↗

Cluster the gallery to work through visually similar images together. Keep box and class rules consistent within a group, and use Outliers first to find unusual images worth checking.

Use Needs review or Model predictions to inspect to focus each pass. Clustering organizes inspection; every image still needs your review.

Cluster your gallery →
Annotension gallery showing Sample Objects and its image review status.
Click an image to open the editor. Command-G returns here.

Current tools and what comes next

Today’s workflow uses axis-aligned bounding boxes. Segmentation and other annotation tools are planned; tool types, formats, and release dates are still undecided.

Follow the local training and review loop. Dataset import/export and clustering support the same workflow.

  • Bounding-box annotationAvailable
  • YOLO / COCO detection import and exportAvailable
  • Local image clusteringAvailable
  • Detector training and suggestion reviewOptional setup
  • Segmentation and other annotation toolsPlanned

Before you start

Mac requirements and getting the app

The app requires macOS 13 or later. Its hardware target is Apple Silicon M2 with 16 GB RAM or above; this is not a performance guarantee for every dataset or training workload.

Install Annotension from its DMG; no Xcode installation is needed. See installation and release status. A public download has not been published yet.

Annotation and clustering run locally without a third-party runtime. Optional detector training requires Apple Silicon, macOS 14 or later, Python 3.11, a separate runtime, and an initial model download. No dataset upload is required.

How projects and reviewed labels work

Creating a project copies your images; the originals stay unchanged. Move or back up the complete project folder, including its images.

Image review is an explicit decision. Inspect an empty image and mark it reviewed to make it a verified negative. Editing boxes clears review.

Accept, edit, or reject model suggestions, then inspect the whole image for missed objects before approving it. Model suggestions do not automatically become reviewed labels.