Auto-label images with local ONNX models

Auto-labeling has two workflows. Interactive segmentation turns point and box prompts into a mask. Automatic models find every supported object in an image in one run and return editable boxes or polygons. Both execute locally through the same bounded ONNX inference core. Predictions remain editable and follow the workspace's normal manual-save or auto-save setting.

The three short recordings below show the interactive workflow with a real construction-safety photograph. Automatic detection and instance segmentation use the same model picker and annotation canvas.

Step 1: Turn on AI and choose a model

Open an Object Detection, Image Segmentation, or Instance Segmentation project, start labelling, then select the brain-shaped AI tool in the left toolbar. The task-aware picker shows only compatible models and separates them into project models, interactive segmentation, object detection, and instance segmentation.

The built-in catalog contains 13 immutable model entries:

• Interactive segmentation: MobileSAM; SAM 2 Tiny, Small, Base+, and Large; SAM 2.1 Tiny, Small, Base+, and Large; and EfficientViT-SAM L0.
• Automatic object detection: D-FINE N and RF-DETR Nano, using their COCO label spaces.
• Automatic instance segmentation: RF-DETR Segmentation Nano, producing editable polygons or rectangles.

MobileSAM and SAM 2 Tiny/Small ship with the installer. Other entries show their download size in the picker and are fetched only when selected. Every archive and declared member is verified before ONNX Runtime loads it. A model already stored locally is marked Ready.

Step 2A: Prompt and refine an object

Start with Include box when the object has a clear extent. Drag a box around the object, then use Include point on a missing region or Exclude point on unwanted background. Every new prompt updates the preview; in this example, a single exclusion point removes the worker's temple while keeping the helmet rim.

The prompt controls also have keyboard shortcuts:

• Include point: A
• Exclude point: D
• Include box: R
• Clear prompts: C
• Finish object: F

Image Segmentation projects can keep the generated Polygon or reduce it to a Rectangle. Instance Segmentation projects can do the same. Object Detection projects always produce a rectangle because their annotations and trained models use bounding boxes. Clear removes the current prompts and preview without deleting labels you already finished.

Step 2B: Run an automatic model

Choose an automatic detector or segmenter, wait for Ready, then select Run model. Predictions are added as editable proposals. Clear predictions removes proposals from that model without deleting manual shapes or accepted predictions from another model.

The catalog detectors use COCO class names and request only classes whose names exist exactly in the project label list. If none match, the app asks you to add a supported label or choose an interactive model. This prevents a prediction from being silently assigned to the wrong project class.

Step 3: Assign a class and save

For an interactive preview, select Finish object, choose its class, and save the annotations. Automatic predictions already carry their matching class, but remain editable before saving. Generated polygons and boxes use the same class color and name as manually drawn shapes. Inspect every proposal before moving to the next image; auto-labeling accelerates annotation, but it does not replace review.

Reuse a model trained in the project

Successful RF-DETR detection and instance-segmentation training runs export an ONNX graph. AnyLearning verifies that artifact and lists it under Project models for the same project, preserving the project's class order. Select it and use Run model like any other automatic model; no manual import or class remapping is required.

Import a YOLO-family ONNX model

In the desktop app, choose Import ONNX beside the model picker. Import accepts a single-file detection or instance-segmentation graph up to 20 GiB and supports YOLOv5, YOLOv8, YOLOv9, YOLOv10, YOLO11, YOLO12, YOLO26, and YOLOX detection profiles. You can use the project labels, the COCO 80 order, or enter an explicit class list in the model's exact output order.

Use Auto-detect unless you know the exporter profile. Set an input size only for a dynamic-input graph. External-data ONNX bundles are not imported through this dialog yet. The selected file is copied into AnyLearning's managed data directory and always runs through the same bounded inference core as catalog models.

First-run behavior

The demonstration image comes from the Safety Helmet and Reflective Jacket dataset and is used under the Apache-2.0 license. The recording shows SAM 2 Hiera-Small's real output in AnyLearning; the box comes from the inferred mask and is not hand-drawn or composited for the guide.