Recognition

assigning semantic labels to pixels, regions, or entire images.

For example:

• Image-level: Does this image contain a building? (Answer: Yes/No)
• Object-level: Where is the car? What are the people doing?
• Attribute-level: What material is the building made of? What time does the clock show? Recognition tasks vary in granularity:
• Category-level: Detecting any cereal box.
• Instance-level: Identifying a specific cereal box (e.g., Kellogg’s).

Challenges

• scale of categories
  • defining categories is also subjective
• variations
  • viewpoint
  • illumination
  • scale
  • deformation (non-rigid objects)
  • occlusion
  • clutter

recognition pipeline

• feature extraction
• train a classifier - learn function f that maps features to labels
• testing: apply f to new, unseen images

feature representation

• color histograms - simple, not robust to scale

• SIFT - local detector - scale-invariant, good for keypoints

• deep learning - highly flexible but data heavy

• k-nearest neighbors (KNN)

how to choose features?

(i think rgb is scale invariant tho)