Learning Embeddings for Indexing and Recognition
Speaker: Vassilis Athitsos , CS Dept., Boston University
Date: March 3 2004
Time: 2:00PM to 3:00PM
Location: Open area in Vision Lab, 6th floor, 400 Tech Sq
Host: Gregory Shakhnarovich, CSAIL
Contact: Gregory Shakhnarovich, gregory@ai.mit.edu
Relevant URL: ABSTRACT:
An embedding method is presented, that can be used to speed up nearest
neighbor retrieval, or to learn a good distance measure for nearest
neighbor classification. Given a space X with a computationally
expensive distance measure (like edge images with the chamfer
distance, or time series with dynamic time warping), we construct an
embedding that maps database and query objects into a Euclidean space,
in which similarities can be rapidly measured using a weighted
Manhattan distance. Embedding construction is formulated as a machine
learning task, where AdaBoost is used to combine many simple, 1D
embeddings into a multidimensional embedding that preserves a
significant amount of the proximity structure in the original
space. An extension of this framework is also described, which can be
used to combine multiple distance measures defined on the original
space X into a single distance measure, that is optimized for
k-nearest neighbor classification. Performance is evaluated using
datasets from various domains, including a hand pose database, a
dynamic gesture database, and several datasets from the UCI
repository.
This is joint work with Jonathan Alon, Stan Sclaroff and George Kollios.
See other events that are part of MIT Machine Vision Colloquium Spring 2004
See other events happening in March 2004
