Please post your thought on the earth mover's distance here.

A simple question:
Can anyone explain why using the lower bound reduces the number of EMD computations? (First paragraph in P9)

EMD can be used to search a large database of images based on a query image. Some of the images will be so different that even the easily computed lower bound will be high. These images can be rejected immediately, and don't have to be checked by a full computation of the EMD.

EMD (and histogram comparison in general) is an interesting approach because it seems to operate on a certain superficial level of "image similiarity". Both, color-based and texture-based EMD methods could end up finding images which have very similiar histograms, but really have fundamentally different meanings (such as p14, fig 7, picture a3). I guess its up to the end user to decide what kind of similarity he is really looking for, superficial appearance (such as images with a similar color-signature), or something deeper, more meaning-based (such as "all images that have the same kind of people/objects performing the same kind of actions under the same kind of environment").

EMD. It seems to me that if the signature of the image lies in a high dimensional space, there is no clear advantage on using
this metric, because the performance benefit is low and it is expensive computationally.
I would like to have seen its performance evaluated on different databases to make sure that the results are not biased.

Hello, Your site is great. Regards, Valintino Guxxi

racorvardomt