Skip to main content
eScholarship
Open Access Publications from the University of California

Multiclass Boosting for Fast Multiclass Object Detection /

  • Author(s): Saberian, Mohammad
  • et al.
Abstract

In this thesis the problem of designing a fast multiclass object detector based on cascade architecture is considered. A classifier cascade is a sequence of simple to complex sub-classifiers where each stage either rejects the input or pass it to the next stage. Since most of the non-target inputs get rejected with the simple sub- classifiers in the early stages of the cascade, the overall classification will be fast. Since cascade sub- classifier are usually trained with Boosting algorithms, the thesis starts with proposing TaylorBoost which explains Boosting algorithms as iterative descent algorithms for minimizing Taylor expansion of risk of classification in function space. In the rest of this thesis TaylorBoost is used to derive appropriate Boosting algorithms based on the requirements of the problems. The main challenge in designing a classifier cascade is to tune speed-accuracy trade-off, e.g. more complex early stages in the cascade may increase accuracy but degrades speed of classification significantly. To address this issue, this thesis proposes a new Boosting algorithm, FCBoost, for designing a classifier cascade by minimizing a Lagrangian risk that jointly accounts for classification accuracy and speed of classification. While FCBoost enables designing cascade detectors for single class of objects, designing detectors for multiple objects is still problematic. This is because each of the object detectors has to be trained independently which results in many redundant computational complexity. To address this issue, the thesis next proposes a new multiclass Boosting framework, MCBoost. Combining FCBoost and MCBoost, makes it possible to learn detector cascade for detecting multiple objects. The remaining challenge is that in a multiclass cascade, early stages should implement binary target vs. non-target detectors of high simplicity and false-positive rate, and late stages should be multiclass classifiers of high accuracy and complexity to distinguish between target classes. The thesis proposes a method to manipulate cost of classification in MCBoost based on cascade false-positive rate to address this issue. Experiments on the problems of multi-view car detection and simultaneous detection of multiple traffic signs show that the proposed detector is faster and more accurate than previous approaches

Main Content
Current View