Abstract:

In this paper, a conditional model (CM) is used to incorporate different feature potentials including texture, texture environment and location features of objects for multiclass object recognition and segmentation in complex natural images. Besides, we model the relationship between different objects by the scene of images and propose a new scenebased conditional model called the sCM model. We investigate the performance of our model in the classbased pixelwise segmentation of images on the Oliva & Torralba database and compare its result with other methods. The results show that our themebased RCRF model significantly improves the accuracy of objects in the whole database. More significantly, a large perceptual improvement is gained, i.e. the details of different objects are correctly labeled.

Key words: object recognition；multiclass；image segmentation；Jointboost