AMMAI reading and notes

pros:
  classification methodologies can be directly applied
  training data are easy to get
cons:
  minimize error in classification rather than ranking
  pairwise computation cost

training image from google, testing on Caltech dataset

unsupervised learning algorithm

pLSA
  no spatial information
ABS-pLSA
  latent variable x (grid centroid coordinate)
  grid methodology, still translation and scaling variant
TSI-pLSA
  latent variable c:(centroid, xscale, yscale)
  1. apply to pLSA
  2. fit a mixture of Gaussian with k=(1, 2, ...K) components to the location of the region, weighted by p(w|z)
  3.