Annotation Best Practices

6. Annotation Best Practices#

  • Label consistency.

    • All instances of all classes in all images must be labelled.

    • Partial labelling will impact performance.

  • Label accuracy.

    • Ensure tight and consistent bounding boxes.

    • Labels must closely enclose each object.

    • No space should exist between an object, and it’s bounding box.

    • No objects should be missing a label.

7. Precision & Accuracy#

  • Precision indicates the consistency of annotations

  • Accuracy refers to how close the annotated data is to the true values

7.1. Example: Initially looks the same#

7.2. Zoom to see the difference in Accuracy#

High accuracy and precision lead to better-trained models by reducing noise

7.3. What about Partially Visible (Occluded) Objects (like urchins)?#

  • Two main approaches. No right or wrong answer.

  • Experimentation and then consistency is key.

  1. Estimate and draw full bounding boxes (for full object learning)

    • pro: helpful for other downstream opencv tasks like segmenation and sizing techniques

    • con: subjective and time consuming.

    • Suggestion: If full object learning is important, try instance segmentation instead

  2. Or Draw Bounding box only around the visible portion (recommeneded)

    • pro: grounded in what is actually visable

    • cons: inconsistent box sizes. postential for more false postives/negatives

    • Important Dataset Note:

      • For this techinque, ensure a good balance of fully visible and partially occluded objects in final training dataset

7.3.1. Example Annotation Flow for Occluded Objects:#

  • Fully visible:

    • Draw a tight bounding box around the entire object.

  • Partially Visable:

    • Draw a box around just the exposed portion

  • Mostly covered:

    • If confident annotate it.

    • But if unsure, just skip it.

Contents