Setup & Annotation Tab

Setup & Annotation is where you define the project structure, label schema, and dataset layout for later training and inference.

At a glance

Best for	Typical output	Usually next
Project setup, annotation launch, dataset preparation	Project scaffold, labels, train/val split, dataset.yaml	Model Training

Main workflow

Set project root
  -> Define keypoints / behaviors / skeleton
  -> Annotate
  -> Create train/val split
  -> Generate dataset.yaml
  -> Run Dataset QA

1. Project root and folder scaffold

Selecting a project root helps IntegraPose keep paths consistent.

Typical structure:

project_root/
  images_all/
  labels_all/
  images/train/
  images/val/
  labels/train/
  labels/val/
  models/
  videos/

2. Define the schema

Field	What to enter
Keypoint names	The exact keypoint order used by your pose project
Behaviors	Optional class IDs and names when your workflow uses behavior classes
Skeleton connections	Optional edges for overlays and annotation visuals

3. Choose a labeling route

Built-in annotator

Use the built-in annotator when you want fully manual labeling from the main app.

Typical setup:

• Image Directory -> your flat image folder, often images_all/
• Annotation Output Dir -> your label folder, often labels_all/

Assisted Pose Curation

Use Assisted Pose Curation when you want review-first pose labeling with model-assisted suggestions.

Typical flow:

1. Enable the plugin if needed
2. Open Open Assisted Pose Curation...
3. Pull candidate frames
4. Review and correct suggested poses
5. Export back into the standard training workflow

4. Create the train/val split

Use Create Train/Val Split Folders after labeling.

This populates:

• images/train
• images/val
• labels/train
• labels/val

The shipped split controls are:

Control	Meaning
Validation split (%)	Percentage of images assigned to validation
Seed	Makes the split reproducible
Move files (don't copy)	Moves source images and labels instead of preserving them; off by default
Include unlabeled images	Includes unmatched images with empty label files when intentional negatives are required
auto strategy	Preserves source-style filename groups when recognizable, otherwise falls back safely
random strategy	Splits individual images without source-prefix grouping
prefix strategy	Always groups files by the text before the selected delimiter

Use auto or prefix for frames flattened by Tab 1 when neighboring frames from one source must not be divided between training and validation. A successful split writes split_manifest.json in the dataset root.

5. Generate dataset.yaml

Use Generate dataset.yaml after the split exists, or point Setup to an existing Ultralytics-style dataset layout.

This file is what the Training tab uses later.

6. Run Dataset QA

Use Dataset QA before training to catch:

• missing files
• malformed labels
• mismatched image/label pairs
• dataset structure problems
• class or keypoint schema problems
• normalized boxes below the configured tiny-area threshold

The QA panel shows each check, severity, count, and details. Use Export QA Report... to preserve the result. Ignore Selected Check (Project) records an intentional project-level waiver; use it only after confirming that the finding is expected. Changing dataset paths or schema marks the previous QA result stale.

Practical tips

• Keep keypoint order stable once training starts.
• Save the project after major setup changes.
• Use Apply Keypoints / Refresh after changing the comma-separated keypoint list so skeleton controls stay synchronized.
• The split and YAML buttons remain unavailable until the required paths are ready; use the Dataset readiness message and Refresh to see what is missing.
• If you are using a detection-only model workflow, Setup may still be useful for project organization, but the built-in Training tab remains pose-oriented.