1. Image Retrieval
Longin Jan Latecki

2. Motivation: The limitation of existing shape similarity measures
Traditional shape similarity measures cannot capture the intrinsic property of the shape. We will use a graph-based transductive learning algorithm to tackle this problem.
IDSC result for
the query fly:
Result after
learning:

3. Image Retrieval as Label Propagation
Basic Assumptions and prior knowledge:
is a set of objects and is a query object.
is a similarity function.
Let be be a similarity matrix, which is also called affinity matrix. We define a sequence of labeling functions with and for Following label propagation iterate:
We have only one class that contains only one labeled element being the query. Hence we stop iterating at some t=T.
Thus, we interpret as a set of normalized similarity values to the query:

4. Affinity Matrix
The affinity matrix W is crucial for the proposed approach. Let be a distance matrix computed by some shape distance function. It is transformed to an affinity matrix W by using a Gaussian kernel:
In our experiment, we use an adaptive kernel size based on the mean distance to K-nearest neighborhoods:
where represents the mean distance of the K-nearest neighbor distance of sample and C is an extra parameter.

5. Retrieval Examples