Structured Labels in Random Forests for Semantic Labelling and Object Detection.

摘要: Ensembles of randomized decision trees, known as Random Forests, have become a valuable machine learning tool for addressing many computer vision problems. Despite their popularity, few works tried to exploit contextual and structural information in random forests order improve performance. In this paper, we propose simple effective way integrate forests, which is typically reflected the structured output space complex problems like semantic image labelling. Our paper has several contributions: We show how can be augmented with label used deliver low-level predictions. The task carried out by employing novel split function evaluation criterion that exploits joint distribution observed space. This allows forest learn typical transitions between object classes avoid locally implausible configurations. provide two approaches integrating predictions obtained at local level from into concise, global, our new ideas also Hough-forest framework view exploiting classification performance on detection. Finally, experimental evidence effectiveness approach different tasks: Semantic labelling challenging MSRCv2 CamVid databases, reconstruction occluded handwritten Chinese characters Kaist database pedestrian detection TU Darmstadt databases.