Complex Features
Jump to navigation
Jump to search
Main Issues
Interest operator
- Interest operators are justified by the need to avoid duplication of features: the interest point provides an `anchor' so that we do not re-learn shifted versions of the same features
- Jurie and Triggs argue for random sampling. But this may be due to the fact that common interest operators select for corners and points and forget edges.
- Interest operators that select for points as well as for edges seem promising.
Shifts
How do we avoid shifts?
- Interest operators
- Compex cell behavior: given a feature, measure distance from an image patch at a given location by computing the maximum of correlation (normalized correlation) in a neighborhood of the assigned location
Duplication
How do we avoid duplication of features?
- interest operator
- shifts
- learn metrics to make `related' features stick together (see Hinton's and Roweis' papers)
Invariance
Learn correct metric (see hinton and roweis) using proper exemplars of what is `same' and what is `different':
- Should use images of same scene taken under different lighting conditions to provide exemplars of same physical patch under different conditions
- Same, use Moreels' database of features to get multiple viewpoints.
Stats to look at
- Histogram of n. of patches that are captured by each visual word
- Compare distance of patches from each center with distance of other centers from that center
- Structure of each set of patches: how many principal components needed to describe the group?
Literature
Using bags of visual words
- "Representing and Recognizing the Visual Appearance of Materials using Three-Dimensional Textons". Thomas Leung and Jitendra Malik. International Journal of Computer Vision,43(1), 29-44, June 2001. PDF (Earlier in ICCV99]
- Weber, Welling and Perona ECCV2000 - Features used in constellation model
- Vidal-Naquet and Ullman - Nature Neuroscience 02 -- Fragments and mutual information
- Dorko and Schmid - 2003 (?) - Using histograms of visual words
Complex cell behavior
- Weber, Welling and Perona ECCV2000 - idea of shifts
- Brendan J. Frey, Nebojsa Jojic: Fast, Large-Scale Transformation-Invariant Clustering. NIPS 2001 - 721-727
- Rómer Rosales, Kannan Achan, Brendan J. Frey: Learning to cluster using local neighborhood structure. ICML 2004
Random sampling vs interest operators
- FeiFei and Perona, CVPR05
- Frédéric Jurie and Bill Triggs, Creating Efficient Codebooks for Visual Recognition, Int. Conference on Computer Vision , 2005 PDF
Other
- Salakhutdinov, R. R. and Hinton, G. E. Learning a Nonlinear Embedding by Preserving Class Neighbourhood Structure. PDF