Learning Image Descriptors with the Boosting-Trick: VCL Lunch Talk

Seminar | November 30 | 12-1 p.m. | Soda Hall, Visual Computing Lab - 510 Soda

Speaker: Mario Christoudias, EECS/UC Berkeley

Sponsor: Electrical Engineering and Computer Sciences (EECS)

Representing salient image regions in a way that is invariant to unwanted image transformations is a crucial computer vision task. In this talk I will present work on learning an invariant image description jointly optimized over both the descriptor weighting and pooling configuration with boosting. Similar to the kernel-trick, the resulting boosting-trick results in a non-linear transformation of the image patch, however, as I will show can be used to define similarity functions better suited for local patch description. We employ gradient-based weak learners that yields a learned descriptor that closely resembles the well-known SIFT. The resulting descriptor can be learned directly from raw intensity patches achieving state-of-the-art performance. I will present results on a benchmark local patch dataset that exhibits large variations in lighting and viewpoint demonstrating the effectiveness of the proposed approach.

Joint work with Tomasz Trzcinski, Vincent Lepetit and Pascal Fua.

Event Contact: 510-643-2614