A zoo and garden of active basis models

Learning with unknown rotations plus unknown locations and scales

In the following experiments, we allow the change of orientations of the objects, sometimes also the right-left flips, in addition to the change of locations and scales as in experiment 5. The training images are cherry picked from internet. We are rather easy on the code in terms of scale change, mainly for the consideration of speed. We tuned two parameters: One is the number of basis elements in the template. The other is the common resize factor applied to all the training images. Other parameters are kept unchanged.

The learning algorithm is initialized from learning a template from the first training image. In a more thorough implementation, we can initialize the algorithm multiple times from cropped image patches such as in Experiment 5.4 as well as Experiment 15. In the current implementation, we assume there is one object in each image, and in each iteration of learning, we crop a maximum likelihood image patch from each image, to be used to re-learn the template. We do not have to assume there is one object in each image, and we may simply keep all the image patches of high likelihoods, and re-learn the template from these image patches weighted by their likelihoods.

Back to active basis homepage