Multi-scale Conditional Random Fields for over-segmented irregular 3D point clouds classification

In this paper, we propose using multi-scale Conditional Random Fields to classes 3D outdoor terrestrial laser scanned data. We improved Lim and Suter’s methods [1] by introducing regional edge potentials in addition to the local edge and node potentials in the multi-scale Conditional Random Fields, and only a relatively small amount of increment in the computation time is required to achieve the improved recognition rate. In the model, the raw data points are over-segmented into an improved mid-level representation, “super-voxels”. Local and regional features are then extracted from the super-voxel and parameters learnt by the multi-scale Conditional Random Fields. The classification accuracy is improved by 5% to 10% with our proposed model compared to labeling with Conditional Random Fields in [1]. The overall computation time by labeling the super-voxels instead of individual points is lower than the previous 3D data labeling approaches.