Module 04 Activity

Scenario

Establish where classical segmentation stops working for your images, before committing to it or to a model.

What you build

An honest assessment of whether your task has a consistent describable appearance, measured against hand-drawn masks.

Steps

1. Run an edge detector and identify three edges that are not object boundaries - a shadow, a reflection, a printed line.
2. Find contours on a mask where two objects touch. Report how many contours were found and how many objects there are.
3. Compare global, Otsu and adaptive thresholding on an unevenly lit image.
4. Assess your task against the assumption: does the thing you want have a consistent describable appearance? Answer in writing.
5. Hand-draw masks for ten images and measure IoU, precision and recall for your segmentation.
6. Say which direction the errors go and what that implies for the fix.

Evidence to hand in

• Three edges that are not object boundaries.
• The touching-objects contour count.
• The three thresholding methods compared.
• Your written assessment of the appearance assumption.
• IoU, precision and recall against ten hand-drawn masks.
• The error direction and the implied fix.

Review checklist

• Non-boundary edges are identified, not just detected.
• The touching-objects case is shown to merge.
• Otsu is shown to fail the same way a global threshold does.
• The appearance assumption is answered honestly, including 'no'.
• Precision and recall are reported alongside IoU.

Mark complete