Higher-order Class-specific Priors For Semantic Segmentation Of 3D Outdoor Scenes

Semantic Segmentation of 3D outdoor scenes involves segmenting and classify-ing the point clouds.Effective Semantic Segmentation can make the computers un-derstand the outdoor scenes and greatly help many applications,such as virtual reality,autonomous driver,city modeling.One direction of semantic segmentation study is to segment point clouds into objects before classification.However,the errors in seg-mentation are inevitably propagated to the classification results.The other direction is to utilize Conditional Random Field for semantic segmentation,and make use of prior information.However,these approaches either rely only on pairwise terms,or base on a single and simple prior information and can't describe the complex interactions in the scenes.We propose a higher-order Conditional Random Field model based on class-specific priors information for semantic segmentation of 3D outdoor scenes.The model can analyse the prior information of different object classes,and describe the corresponding class-specific higher-order energy function.Consequently we case this model in an energy minimization framework and propose the detailed unary,pairwise and higher-order energy functions.After describing the energy function,we discuss the energy minimization prob-lem.Traditional,pairwise energy function has effective optimal methods,such as graph cuts and loopy belief propagation.However,it's challenging to minimize higher-order energy function.Then the SOSPD algorithm is adapted to optimize the energy function in our method,and finally give the semantic segmentation results.To evaluate the performance of our method,we provide both quantitative and qualitative results on a challenging dataset.The results show an average F1-score of 0.82 compared to the state-of-the-art F1-score of 0.73.