Memory-controlled incremental SVM
Barbara Caputo
Centre for Autonomous Systems and Computer Vision
KTH Stockholm, Schweden
Montag, 28.11.2005, 16 Uhr s.t., H9
In this talk I will present a novel SVM-based algorithm
for visual object recognition, capable of learning model
representations incrementally. The technique combines an incremental
extension of SVMs with a method which reduces the number of support
vectors needed to build the decision function without any loss in
performance. The resulting algorithm is guaranteed to achieve the
same recognition performance as the original incremental method while
reducing the memory requirements. The novel technique was benchmarked
against the batch method and the original version of incremental SVM.
Experiments were performed in two domains, material categorization
and indoor place recognition. In both applications, results show
that the two incremental methods preserve
the performance of the batch algorithm, but only the new
technique consistently achieves a
statistically significant reduction of the memory requirements.
I will then present an extension to
the part of the algorithm controlling the number of support vectors
to be stored. It consists of the introduction of
a parameter which permits a user-set trade-off
between performance and memory reduction.
This property is potentially useful in applications like
indoor place recognition for multi-sensory topological mapping,
where the memory size of the visual models must be kept
under control. Although the method is used here within an incremental
scheme, it can be used for any SVM-based classification
algorithm. Results in both domains of material categorization and
place recognition shows that it is possible to achieve a consistent
reduction of the memory requirements with only a moderate loss in
performance. For example, experiments show that
when the user accepts a reduction in recognition rate of 5%, this
yields a memory reduction of up to 50%.