Experiments in Binning Local Image Statistics

1
Experiments in Binning Local Image Statistics

• Nimar S. Arora

2
Local Image Statistics

• Magnitude of responses to a series of filters.
• Used for matching corresponding points in two
  images for object recognition
• Distribution of local image statistics over the
  image can be used for image categorization as
  well.
• Problem Too many numbers of varying magnitude.
• Usual solution is too aggregate them in bins --
  spatial bins and magnitude bins.
• My work is to study these bins

3
Evaluation of binning

• CalTech 101 dataset
• Compute histograms of brightness and gradient
  values at each pixel for all the images
• Classify a test image by nearest neighbour in the
  training image using Chi-square distance

4
Spatial Binning Log-polar vs Rectangular Grid
80
69
10 categories with 10 training and 10 test images
(randomly chosen)
5
Brightness log bins vs linear bins
80
72
10 categories with 10 training and 10 test images
(randomly chosen)
6
Gradient logbins vs linear bins
80
78
10 categories with 10 training and 10 test images
(randomly chosen)
7
Other negative results

• Bins based on relative distances are worse (66)
  than bins based on absolute distances
• Smoothing the image prior to computing gradients
  doesnt help in general (very few exceptions)

8
Conclusions

• Use log-polar grids
• Bin the brightness on a log-scale
• Bin the gradients on a linear scale
• My current numbers on CalTech 101 using this
  simple technique 30 (with 30 training images
  and at most 30 test images from all categories)