Learning Local Image Descriptors
By Matthew Brown
This talk will discuss methods to learn local image descriptors using calibrated
imagery as training data. These descriptors will be useful for image matching
and object recognition, and I will discuss applications in panoramic stitching,
3D reconstruction and image search.
In contrast to previous approaches, that have required manual tuning of many
parameters and resulted in high-dimensional descriptor vectors, we use
discriminative techniques to learn optimal low-dimensional descriptors. We
propose 2 algorithms for this purpose: 1) direct parameter tuning using Powell
Minimisation, and 2) discriminative dimensionality reduction using Linear
Discriminant Analysis (LDA).
Our new descriptors will be shown to outperform the state-of-the-art on our test
data, as well as giving interesting insights into successful existing designs.
Bio:
Matthew is a Postdoctoral Fellow (as of Jan 08) at the University of British
Columbia. Before that he was at the Vision Technology Group at Microsoft
Research in Redmond, where he worked on image matching methods for panoramic
stitching and 3D reconstruction. He obtained the PhD degree from UBC in 2005.