A Novel Learned Model for Boundary Detection
Boundary detection is a critical, well-studied computer vision problem. Clearly it would be nice to have algorithms which know where one object stops and another starts. Traditional approaches look for intensity discontinuities in an image, however we believe that detecting boundary from a single image is fundamentally difficult, whereas machine learning techniques have a promising prospect on boundary detection. A novel learning model is proposed in this paper. Random decision forest classifier trained by human segmentation images will applied to determine the mapping from image patches to boundary probabilities. The results on BSDS500 dataset show that our model ties the performance of other competing approaches, but being magnitude faster than others.