Research Group Leader

Prof. Dr. Matthias Bethge
Prof. Dr. Matthias Bethge
Phone: +49 7071 29-89017
Fax: +49 7071 29-25015
mbethge[at]tuebingen.mpg.de

 

Secretary: Heike König
Phone: +49 7071 29-89018
Fax: +49 7071 29-25015
heike.koenig[at]tuebingen.mpg.de
 

 

Hierarchical Modeling of Local Image Features through Lp-Nested Symmetric Distributions

Data
All experiments below have been carried out in ten pairs of training and test sets sampled uniformly from all images of the van Hateren database. The patch size is either 8x8 or 16x16.

The data objects for non-CGC data can be downloaded here: 8x8,16x16


The data objects for CGC data can be downloaded here: 8x8

All experiment scripts below assume that the respective data is located in a subdirectory called "data/", i.e. if the script is located in "./" then the data must be in "./data/". Apart from that you should have a subfolder called "./results/" in the respective directory.

 
 
Software
All experiments below need the NISDET toolbox version 1.1 with all its other requirements (lightspeed and fastfit).
 
Experiments for Table 1 and 16x16 in the text



 
ISA 1
Compute ALL on for ISA models on WO-data.
8x8 WO
Trained Models
ISA 2
Compute ALL on for ISA models on CGC-data.
8x8 CGC
Trained Models
PND 1
Compute ALL on for PND DT model and for PSSD model on WO-data.
8x8 WO
Trained Models
PND 2
Compute ALL of PND ISA-like models on WO-data.
8x8 WO
Trained Models
PND 3
Compute Multi-Information between inner nodes.
8x8 WO ,8x8 CGC
-
PND 4
Compute ALL on DT model for CGC-data.
8x8 CGC
Trained Models
PND 5
Compute ALL of PND ISA-like models on CGC-data.
8x8 CGC
Trained Models
PND 6
Compute ALL of ICA on CGC-data.
8x8 CGC
Trained Models
PND 7
Compute ALL of PND DT model on 16x16 WO data.
16x16 WO
Trained Models
PSSD 1
Compute ALL for PSSD model on WO-data.
16x16 WO
Trained Models
 
Figure 2



Figure 2 can be reproduced by loading the respective models from ISA 1 and PND 2, sampling data from those models using the nisdet toolbox command sample and plotting the Lp-radii of the two subspaces agains each other.
For PND 2:

>> load results/results8_s2_1.mat
>> dat = sampledata(dpnd_mln{1},100);
>> L1 = getSubTree(dpnd_mln{1}.param.f,1);
>> L2 = getSubTree(dpnd_mln{1}.param.f,2);
>> plot(L1(dat),L2(dat),'.') For ISA 2:

>> load results/results8_s2_1
>> dat = sampledata(disa8{1},100);
>> plot(norm(dat(disa8{1}.param.subspaces{1},:),disa8{1}.param.subdist{1}.param.p),...
norm(dat(disa8{1}.param.subspaces{2},:),disa8{1}.param.subdist{2}.param.p),'.')
For the other two subfigures one has to do the same despite using natural image data instead of samples from the distributions.
Last updated: Wednesday, 27.02.2013