Large-scale conditional random field for natural outdoor scene labeling

We propose a novel large-scale conditional random field model with respect to the problem of natural outdoor scene labeling. The novelty of the proposed method lies in three aspects: 1. features from two neighboring regions are concatenated to form the input of the pair-wise classifier to compensate for the simultaneous feature deviation of neighboring regions; 2. the definition of a generalized neighboring system and the incorporation of direction-specific patterns in conditional random field models based on the generalized neighboring system to better simulate the visual cognition of human being; and 3. the definition of a similarity criterion based on the bags-of-words expression to facilitate the incorporation of semantic patterns. The proposed model is first evaluated over the Corel dataset. Both qualitative and quantitative results show that our model is capable of modeling large-scale spatial relationships between objects in natural outdoor scenes, and achieves better results than other existing conditional random field models. Furthermore, our model is also evaluated over several other natural datasets, which are taken from logged field tests, to further demonstrate the adaptability of our model to different lighting conditions.