Semantic segmentation using regions in natural scenes

By introducing an over-segmentation algorithm into the conditional model (CM), we propose a new region-based CM model (R-CM), and investigate its performance on semantic segmentation of images. In order to incorporate structure information of objects, we segment an image into regions by using an over-segmentation algorithm. Based on the results of CM model, we first consider assigning all pixels in one region with the same label, and then other feature potentials are included to counteract the influence of false over-segmentation. We compare our results to related work on the Olive & Torralba database and show that aside from improved accuracy of the whole database, our model obtains a perceptual improvement, with boundary of different objects correctly labeled.