Image retrieval

The problem of image retrieval is stated as follows. Given a query image, and a collection of arbitury images. We want to extract an image from the collection which is perceptually the closest to the query.

Our proposed method is to extract the epitome from the query image, use it as a metric for the retrieval. The retrieved image $I^*$ based on the query image $I_q$ should meet

$$I^* = {\hbox{$\underset{I}{\mbox{argmax}}\;$}} E [ \log p(\{Z_k(I)\}_{k=1}^P, \{T_k\}_{k=1}^P \vert e(I_q))$$ (23)

where $Z_k(I)$ is the collection of patches generated from image $I$, and $e(I_q)$ is the parameter of the epitome of query image $I_q$.

We ran the experiments on a 20-image database, which can be classified into 5 classes by subjective inspection. We picked 5 images, each from one class, and trained their epitomes. For each of these epitome, we retrieve an image from the other 19 images according to the rule stated in 23

The query on the zebra obviously failed using our method. One explaination is that the epitome is not object-oriented. Although it can caputure small patterns, it is unable to interprete an image as a whole.

Test set:

Results: