Segmentation d'images

Semantic segmentation identifies categories, while instance segmentation distinguishes separate objects of the same category. The resulting mask is an estimate whose boundaries and missed regions need evaluation.

Points clés à retenir

• Distinguish semantic and instance tasks.
• Define annotation boundaries.
• Evaluate minority regions and coordinate mapping.

Choose the segmentation task before collecting annotations. Labeling every road pixel is different from identifying each individual vehicle. Define how to treat uncertain boundaries, transparent objects, overlapping instances, and regions outside the label set. Annotation quality affects the result. Two reviewers may draw different boundaries around hair, shadows, or partially visible objects. Document the labeling convention and measure disagreements rather than assuming there is always one perfectly obvious mask. Use metrics suited to the application. Pixel accuracy can look high when most pixels are background. Overlap metrics such as intersection over union can provide more information, but class balance and boundary quality still matter. A small boundary error may be harmless in one task and consequential in another. Test the full image pipeline. Cropping, resizing, and coordinate conversion can shift an otherwise reasonable mask when it is placed back on the original image. Preserve source dimensions and inspect overlays at the scale where the result will be used.

1. Construct an image with 1,000 pixels, of which 950 are background and 50 belong to the target object.
2. A prediction marking every pixel as background has 95% pixel accuracy but detects none of the object.
3. Inspect class-specific overlap and missed-object behavior rather than reporting only the overall pixel score.

The invented pixel counts illustrate an evaluation pitfall.