A discriminatively trained, multiscle, deformable part model
P. Felzenszwald, D. McAllester, D. Ramanan, Proc. CVPR 2008.
Structured learning techniques applied to recognition using parts-based models. Specifically, extending the HoG representation to use parts instead of a single window.
Subscribe to:
Post Comments (Atom)
To answer Martial's question on handling multiple views:
ReplyDeleteIn the latest (journal) version of the work, an object model can be composed/mixture of multiple "components". A component is an instance of the learned filter weights (for root and parts) and distance function parameters; so what I presented today would is a model composed of 1 component. The need for multiple components to represent an object lie exactly when the object looks completely different, e.g. see Figure 2 in
P. Felzenszwalb, R. Girshick, D. McAllester, D. Ramanan
Object Detection with Discriminatively Trained Part Based Models
To appear in the IEEE Transactions on Pattern Analysis and Machine Intelligence
http://people.cs.uchicago.edu/~pff/papers/lsvm-pami.pdf
Where 1 component intuitively models the frontal view of the bike and the other component models the side view.
Detection and learning are straightforward:
For detection, the max score over the components is used. That is, if your model is composed of C components, then given a root window, run the black-box inference C times and use the parts from the component with max score.
For learning, the parameters you optimize over are now the concatenation of the each component's parameters. The subgradient update then modifies the respective component parameters given from the max component score.