search

liboyin / img-classify

Household Object Recognition Challenge
GNU General Public License v2.0
0 stars 0 forks source link

On what level do SIFT descriptors cluster? #1

Open liboyin opened 9 years ago

liboyin commented 9 years ago

SIFT descriptors from visual features that look similar should naturally cluster. It remains to be seen whether SIFT descriptors from the same class would cluster. This hypothesis should be tested by comparing the within-class variance and between-class variance. Also, to see how strongly do SIFT descriptors cluster, we should try to plot the distance from each query descriptor to all training descriptors.

liboyin commented 9 years ago

Libo is working on it. This issue is important for the explanability of the kNN and set nearest neighbour mode.

liboyin commented 9 years ago

In other words:

  1. How separable are all SIFT feature descriptors? (between-descriptors)
  2. How separable are classes? (between-classes & within-class)

and for the kNN/SNN model:

  1. For one test SIFT descriptor from a test image, how separable are classes?
  2. For all test SIFT descriptors from a test image, how separable are classes?
liboyin commented 9 years ago

The last two hypothesis for SNN model (https://github.com/liboyin/horc/issues/2) have been confirmed. As to the first two hypothesis, judging from the heatmap, SIFT feature vectors do cluster, yet classes do not. Actual clustering is required to test these two hypothesis.

liboyin commented 9 years ago

ref: http://scikit-learn.org/stable/modules/clustering.html