Attention-Driven Parts-Based Object Detection
Ilkka Autio and Jussi Lindgren
In: ECAI 2004, 22-27 Aug 2004, Valencia, Spain.
Recent studies have argued that natural vision systems
perform classification by utilizing different mechanisms
depending on the visual input. In this paper we present
a hybrid, data-driven object detection system that combines
parts-based matching and view-based attention for faster detection.
We propose a simple competitive policy that allows incremental
addition of new object classes to the system without requiring
class-vs-class training. Using our framework, we show empirical
support for the hypothesis that low-frequency visual information
can be effectively used to direct attention and possibly subsume
further, more costly analysis. We evaluate our approach on face
and car detection problems, while concentrating on the capability
to learn from small samples. Our implementation
is freely available as Matlab source code.