Appendix Classification - AsyDynamics/CS231n GitHub Wiki

This is the notes about Image Classification, extra material for CS231n course.
Source: http://cs231n.github.io/classification/

Challenge

  1. challenge
  • Viewpoint validation
  • Scale variation
  • Deformation
  • Occlusion - Sometimes only a small portion of an object (as little as few pixels) could be visible
  • Illumination condition
  • Background clutter
  • Intra-class variation - different type
  1. pipeline
  • input - N images, labeled with K class
  • learning - training a classifier, or learning a model.
  • evaluation

Nearest Neighbor Classifier

It's nothing to do with Convolutional Neural Networks.
Dataset CIFAR-10

  1. L1 distance
    $ d=sum(I_1-I_2) $

  2. L2 distance - euclidean distance $ d=sqrt(sum(I1^2-I2^2)) $

k - Nearest Neighbor Classifier

Instead of finding the single closest image in the training set, find the k top closest images and have vote on the label of the test image. Higher K have a smoothing effect that makes the classifier more resistant.

Validation sets for hyper-parameter tuning

  • Validation set
  • Cross-validation