Image Classification: A core task in Computer Vision

• Core task in computer vision
• System
  • receives input image
  • aware of pre-determined set of categories/labels
  • job: look at the picture and assign it to specific category

Semantic Gap

• Difference between how human and computer sees images
• Computer sees image as big grid of numbers
  • 3 channels: RGB

Challenges

• algorithm should be robust to all such challenges
• Viewpoint Variation
  • Small changes will change the pixel grid entirely
    • ex) change of viewpoint
• Illumination
  • Lighting conditions can also change the grid entirely
• Deformation
  • Varied poses and positions
• Occlusion
  • Might not be given all part of object
• Background Clutter
  • Object might be similar to background
• Intraclass Variation
  • ex) cats: different colors, sizes, ...
• No obvious explicit algorithm to make it happen

Attempts that have been made

Find edges, find corners

• find all edges and corners
• write down explicit rules to execute classification
• X work
  • Very brittle
  • Changing category will result in changing everything all over