Revision history [back]

Here, I assume you are using a bag-of-words like approach. Therefore, you have histograms of your descriptors, which seems more efficient in your case. The idea is to build a tree of histograms. You can use different solutions:

• a binary tree on each bin: if the bins is greater than the mean, go left, otherwise go right, and so one
• a binary tree on half(third/...) of histogram's bins
• cluster your histograms with a metric (intersection, chi-square, Bhattacharyya,...) and build a spatial grid according to the neigborhood (KNN).

I think there is many others solutions, but I'm not familiar with image retrieving optimization, try Google it and read some papers... This is only the "first" solutions I will use if I have to do something similar. Let us know the methods you've try!

Enjoy.