Top-down category recognition

Alex, Eugene, Pietro: please add or change whatever you want.

Target

  CVPR 2007 ?

Final Paper Should Cover These Topics

1. Automatic taxonomy creation for any set of categories
   1. We no longer have to make the trees by hand
   2. This enables automatic top-down category searching
2. Efficiency increase by descending tree
   1. You quickly eliminate most of the possibilities
   2. This should increase speed, right?
   3. Demonstrate and quantify this
3. Classification performance
   1. Quantify this too
   2. Does it improve?

Experiment #1 (minimum needed to publish?)

1. Don't build the entire hierarchy: just the first level (4 branches)
   1. Do this in an automated fashion
2. Top-level classifiers
   1. Train on n images from each of these 4 subgroups
      1. What value of n is optimal?
      2. Speed/performance tradeoff
   2. Should there be a 5th branch that says, "I'm not sure"?
3. Bottom-level classifiers
   1. Choose a different set of one-vs-all SVMs for each of the 4 branches
   2. This means we only have to check for categories that are in our branch
   3. Should be about 75% less work!
4. Testing phase
   1. Compare to standard one-vs-all paradigm
      1. Classification performances
      2. Training and Testing Speed

Experiment #2

1. Try trees with 1, 2, 3 and 4 levels
2. Compare speed/performance
3. How many N_train per level?
   1. Again a speed/performance tradeoff
   2. Performance may saturate esp. at higher levels
   3. Do we have to find this emperically?
   4. Or can it be guessed from first principles?