Forest Extension of Error Correcting Output Codes and Boosted Landmarks

Abstract

In this paper, we introduce a robust novel approach for detecting objects category in cluttered scenes by generating boosted contextual descriptors of landmarks. In particular, our method avoids the need of image segmentation, being at the same time invariant to scale, global illumination, occlusions and to small affine transformations. Once detected the object category, we address the problem of multiclass recognition where a battery of classifiers is trained able to capture the shared properties between the object descriptors across classes. A natural way to address the multiclass problem is using the error correcting output codes technique. We extend the ECOC technique proposing a methodology to construct a forest of decision trees that are included in the ECOC framework. We present very promising results on standard databases: UCI database and Caltech database as well as in a real image problem