Stacked Keypoint Evidence Trees Schema

[scottclowe]

Replication of a method developed for the BugID project http://web.engr.oregonstate.edu/~tgd/bugid/

Of interest is in particular is this paper: http://web.engr.oregonstate.edu/~tgd/publications/cvpr2009-evidence-trees.pdf

[scottclowe]

Maybe we should email and ask for a copy of their code.

[scottclowe]

I reread the paper and I'm pretty sure I can implement it.

[scottclowe]

Going to need more issues for this. This issue is to get the main schema up and running.

[scottclowe]

We will need to implement their model hierarchy (see Figure 2), which is trained as follows.

• Partition the training data into Train1 and Train2
• For each detector
  • For each image in Train1
  • Detect all keypoints with this detector & describe them
  • Train a random forest (RF1) The random forest takes the description of a single keypoint as input and outputs the probability of the keypoint being taken from each of the plankton classes We train it on all the keypoints detected in all the images in the dataset Train1
• For each image in Train2
  • For each detector
  • Detect and describe all keypoints
    • With each keypoint, predict the plankton class
  • Pool predictions from all keypoints provided by the detector
  • Stack the pooled predictions from each detector together, as a vector
• Train a second random forest (RF2), with boosting This takes as its input the pooled predictions from each detector from an image and outputs the probability the image belongs to each plankton class

[scottclowe]

NB: in a more recent paper, they do pretty much the same thing, but with SVM instead of RF for the second layer.

http://web.engr.oregonstate.edu/~tgd/publications/WACV2011-stacked-spatial-pyramid.pdf