Experiments for Caltech 101 dataset as reported in the paper:

Max-Margin Additive Classifiers for Detection, ICCV'09

Subhransu Maji (smaji@cs.berkeley.edu) 
Alexander C. Berg (aberg@cs.columbia.edu)


Features:
Our features are based on the "weak features" used for Caltech-101 in:

Beyond Bags of Features: Spatial Pyramid Matching 
for Recognizing Natural Scene Categories, 
Lazenbnik, Schmid and Ponce, CVPR'06


In particular we use two scales for the canny edge detector and concatenate the 
features from both the scales together. Our numbers
are similar to the numbers reported in the original paper. 


Classifiers:
We compare the following classification algorithms on these features
1) Linear SVM on the features
2) Intersection Kernel SVM on the pyramid features. This is same as the 
spatial pyramid match kernel. 
3) \phi_1 | \phi_2 encoding + Linear SVM
4) \phi_2 encoding + piecewise linear encoding 


Files:
doit.m : Basic file for doing 1-vs-1 and 1-vs-all classification
using linear and intersection kernel SVMs.

pwl_doit.m : runs the piecewise linear training using various 
encodings. Uses the phi2 encoding and trains a pwc and pwl model as 
well as a liblinear model.

snow_doit.m : runs the training using phi1 (or snow) encoding.

 

