pre-trained 3D SparseConvnet model for ALS point clouds classification in urban areas

Model_Vienna is a pre-trained 3D SparseConvnet[1] model used for point clouds classification in urban area. The model takes ALS (airborne laser scanning) point clouds as input, conducts a voxel-wise classification and maps the prediction back into each point. The considered classes are ground (2), vegetation (5), building (6) and others (8). Others include all the objects above ground except buildings and vegetation, like street lamp-pole, cars and fences. The setup and performance about the model can be found in the paper[2]. Noe that, beside XYZ, this model need echo number and number of echo as input for inference. The training ALS tiles can be found on ZENODO[3]. Model_Vienna_xyz is also a pre-trained 3D SparseConvnet[1] network for ALS point clouds classification in urban areas. This model only takes XYZ as input for inference, no further attribute is needed. This model uses the same structure as Model_Vienna does. However, Model_Vienna_xyz and Model_Vienna uses different tiles of ALS Vienna dataset for training. Four adjacent tiles in the city center of Vienna are usefd for training fro Model_Vienna_xyz, and the four tiles used for training of Model_Vienna locate in different regions of Vienna, as described in [3]. summary of the Model_Vienna and Model_Vienna_xyz name Model_Vienna Model_Vienna_xyz input attributes XYZ, echo number and number of echo XYZ voxel size for training 8cm 8cm number of layers 7 7 block repetition 2 2 whether use residual block True True library Pytorch Pytorch [1] B.Graham,M.Engelcke,andL.vanderMaaten,“3Dsemanticsegmen- tation with submanifold sparse convolutional networks,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2018, pp. 9224–9232 [2] N. Li, O. Kähler, and N. Pfeifer. "A comparison of deep learning methods for airborne lidar point clouds classification". IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2021. 14, 6467-6486. [3] N. Li, N. Pfeifer, and C. Ressl, “Tiles of airborne laser scanning point clouds of Vienna, Austria (2016)” [Data set]. Zenodo, 2021. [Online]. Available: http://doi.org/10.5281/zenodo.477708