1. The document proposes a semantic octree framework that unifies recognition, reconstruction, and representation of 3D scenes. 2. It combines an associative higher-order random fields model for semantic segmentation with an octree representation used in reconstruction to label voxels with semantic classes. 3. Inference is performed over the octree structure to semantically label the reconstructed 3D scene in a computationally efficient manner by exploiting the hierarchical grouping of space provided by the octree.