Supervised object boundary detection based on structured forests

Abstract

Object boundary detection is an interesting and challenging topic in computer vision. Learning and combining the local, mid-level and high-level information play an important role in most of the recent approaches. However, few characteristics of a certain type of object are exploited. In this paper, we propose a novel supervised machine learning framework for object boundary detection, which makes use of the specific object features, such as boundary shape, directions and intensity. In the learning process, structured forest models are employed to tackle the high dimensional multi-class problem. Various experiment results show that our framework outperforms the competing models in the proposed data set, indicating that our framework is highly effective in modeling boundary for specific type of objects.