Semantic Instance Labeling Leveraging Hierarchical Segmentation

摘要: Most of the approaches for indoor RGBD semantic labeling focus on using pixels or super to train a classifier. In this paper, we implement higher level segmentation hierarchy obtain better training our By focusing meaningful segments that conform more directly objects, regardless size, random forest decision trees as classifier simple features such 3D LAB color histogram, width, height, and shape specified by histogram surface normal's. We test method NYU V2 depth dataset, challenging dataset cluttered environments. Our experiments show achieves state art results both general introduced (floor, structure, furniture, objects) object specific labeling. from yields than pixels, patches, in previous work.