Skip to main content

Best practices

Following these annotation guidelines will help you create accurate, consistent annotations that lead to better model performance.

Selecting the appropriate label type​

Choose the right annotation shape based on your object's geometry and use case.

Match the shape to the object​

Each defect or object should be annotated using the most appropriate geometric representation:

Example: 

Oocysts are annotated with polygons rather than rectangles or circles because their geometry is irregular and cannot be accurately represented by simple bounding shapes.
General rule
  • Regular shapes (circles, squares) → Use bounding boxes
  • Irregular shapes → Use polygons
  • Complex contours → Use detailed polygonal annotation

Localization: Polygons vs. bounding boxes​

Choose your localization method based on whether shape information matters.

Decision framework​

Use bounding boxes when:

  • ✅ The precise contour is not relevant
  • ✅ Approximate location is sufficient
  • ✅ Objects are roughly rectangular
  • ✅ Speed of annotation matters

Use polygons when:

  • ✅ Object shape is important
  • ✅ Contour information must be preserved
  • ✅ Directionality or orientation matters
  • ✅ Irregular shapes need accurate representation
Traktech case study

In Traktech's bottle orientation detection, polygonal labeling is required to determine directionality. A simple bounding box cannot capture whether a bottle is upright or tilted, making polygon annotation essential for this use case.

Defining which elements should be labeled​

Establish clear labeling criteria before starting your annotation work and communicate them to all team members.

Create a labeling policy​

A clear labeling policy must be established to determine which objects, structures, or defects are relevant to the task and therefore require annotation.

Key considerations:

  • What objects are critical to your detection task?
  • What can be safely ignored?
  • What are the edge cases?
  • How should ambiguous cases be handled?
Consistency is key

All annotators must follow the same policy to ensure dataset quality and model reliability.

Separating categories into distinct labels​

Organize your labels strategically from the beginning.

Benefits of label separation​

When defects are categorized into separate labels from the outset, it allows more flexibility during model training.

Advantages:

  1. Selective training - Include or exclude categories as needed
  2. Easy experimentation - Test different label combinations
  3. Simpler updates - Add new categories without relabeling
  4. Better analytics - Track performance per category

Example workflow:

❌ Bad: Single label "defect" for all issues
✅ Good: Separate labels "scratch", "dent", "discoloration"

This way, specific categories can later be excluded by simply omitting their labels when configuring the dataset.

Quick guidelines summary​

  • Choose annotation shapes that match object geometry
  • Document a clear labeling policy before starting
  • Prefer segmentation over classification when possible
  • Use bounding boxes for simple localization, polygons for shape-critical tasks
  • Separate different defect types into distinct labels
  • Ensure all annotators follow the same guidelines
  • Review edge cases and document decisions