Performing semantic segmentation of 3D data using deep learning
View Patent ↗A deep artificial neural network (DNN) for generating a semantically-segmented three-dimensional (3D) point cloud is manufactured by a process including obtaining a 3D point cloud, establishing a DNN topology, training the DNN to output labels by subdividing the point cloud, pre-processing the subdivisions, updating weights, and storing weights. Training a DNN includes obtaining a 3D point cloud, establishing a topology of the DNN, training the DNN to output point labels by subdividing, pre-processing the subdivisions, analyzing the features and respective labels of the point cloud to update DNN weights, and storing the weights. A server includes a processor and a memory storing instructions that, when executed by the processor, cause the server to obtain a 3D point cloud, establish a DNN topology, train the DNN to output labels by subdividing, pre-process the subdivisions, analyze the features and respective labels of the point cloud to update weights, and store the weights.
1. A deep artificial neural network for generating a semantically-segmented three-dimensional point cloud stored on a computer readable storage medium, wherein the deep artificial neural network is manufactured by a process comprising:
obtaining one or more training three-dimensional point clouds, wherein one or more features of the point clouds include a respective label, and wherein at least one of the features corresponds to an outdoor structure;
establishing a topology of the deep artificial neural network, wherein the network topology includes an input layer and an output layer having a predetermined number of neurons,
training the deep artificial neural network to output a set of point labels corresponding to an outdoor scene by, for each of the training three-dimensional point clouds:
subdividing the each point cloud into a plurality of subdivisions,
preprocessing the plurality of subdivisions, and
analyzing the features and respective labels of the training three-dimensional point clouds to update a set of weights of the deep artificial neural network; and
storing the set of weights of the deep artificial neural network on the computer readable storage medium as parameters for initializing the deep artificial neural network.
2. The deep artificial neural network of claim 1 , further comprising, wherein the pre-processing includes:
computing a distribution of points across the plurality of subdivisions,
one or both of (i) downsampling a set of points within the plurality of subdivisions, and (ii) upsampling a set of points within the plurality of subdivisions,
spatially normalizing a set of points within the plurality of subdivisions, and
collapsing a set of points within the plurality of subdivisions onto a unit sphere.
3. The deep artificial neural network of claim 2 , further comprising:
generating the semantically-segmented three-dimensional point cloud by using the trained deep artificial neural network to analyze a three-dimensional point cloud not used for training the deep artificial neural network, and
storing the semantically-segmented three-dimensional point cloud on the computer-readable storage medium, wherein the point cloud comprises the set of point labels.
4. The deep artificial neural network of claim 2 , wherein one or more of the point labels is determined using a pairwise distance function.
5. The deep artificial neural network of claim 2 , wherein one or more of the point labels is determined by selecting an arg max of a vector of labels corresponding to the one or more point labels.
6. The deep artificial neural network of claim 1 , wherein each of the plurality of subdivisions is columnar in shape.
7. The deep artificial neural network of claim 1 , wherein the set of point labels includes a type and an elevation.
8. A computer-implemented method for training a deep artificial neural network to generate a semantically-segmented three-dimensional point cloud, comprising:
obtaining one or more training three-dimensional point clouds, wherein one or more features of the point clouds include a respective label, and wherein at least one of the features corresponds to an outdoor structure;
establishing a topology of the deep artificial neural network, wherein the network topology includes an input layer and an output layer having a predetermined number of neurons;
training the deep artificial neural network to output a set of point labels corresponding to an outdoor scene by, for each of the training three-dimensional point clouds:
subdividing the each point cloud into a plurality of subdivisions,
preprocessing the plurality of subdivisions, and
analyzing the features and respective labels of the training three-dimensional point clouds to update a set of weights of the deep artificial neural network; and
storing the set of weights of the deep artificial neural network as parameters for initializing the deep artificial neural network.
9. The computer-implemented method of claim 8 , wherein the pre-processing includes:
computing a distribution of points across the plurality of subdivisions,
one or both of (i) downsampling a set of points within the plurality of subdivisions, and (ii) upsampling a set of points within the plurality of subdivisions,
spatially normalizing a set of points within the plurality of subdivisions, and
collapsing a set of points within the plurality of subdivisions onto a unit sphere.
10. The computer-implemented method of claim 9 , further comprising: generating the semantically-segmented three-dimensional point cloud by using the trained deep artificial neural network to analyze a three-dimensional point cloud not used for training the deep artificial neural network, and
storing the semantically-segmented three-dimensional point cloud on the computer-readable storage medium, wherein the point cloud comprises the set of point labels.
11. The computer-implemented method of claim 9 , wherein one or more of the point labels is determined using a pairwise distance function.
12. The computer-implemented method of claim 9 , wherein one or more of the point labels is determined by selecting an arg max of a vector of labels corresponding to the one or more point labels.
13. The computer-implemented method of claim 8 , wherein each of the plurality of subdivisions is columnar in shape.
14. The computer-implemented method of claim 8 , wherein the set of point labels includes a type and an elevation.
15. A server comprising: one or more processors; and a memory storing instructions that, when executed by the one or more processors, cause the server to
obtain one or more training three-dimensional point clouds, wherein one or more features of the point clouds include a respective label, and wherein at least one of the features corresponds to an outdoor structure;
establish a topology of the deep artificial neural network, wherein the network topology includes an input layer and an output layer having a predetermined number of neurons,
train the deep artificial neural network to output a set of point labels corresponding to an outdoor scene by, for each of the training three-dimensional point clouds:
subdividing the each point cloud into a plurality of subdivisions,
preprocessing the plurality of subdivisions, and
analyzing the features and respective labels of the training three-dimensional point clouds to update a set of weights of the deep artificial neural network; and
store the set of weights of the deep artificial neural network on the memory as parameters for initializing the deep artificial neural network.
16. The server of claim 15 , the memory storing further instructions that, when executed by the one or more processors, cause the server to
compute a distribution of points across the plurality of subdivisions,
one or both of (i) downsample a set of points within the plurality of subdivisions, and (ii) upsample a set of points within the plurality of subdivisions,
spatially normalize a set of points within the plurality of subdivisions, and
collapse a set of points within the plurality of subdivisions onto a unit sphere.
17. The server of claim 16 , the memory storing further instructions that, when executed by the one or more processors, cause the server to
generate the semantically-segmented three-dimensional point cloud by using the trained deep artificial neural network to analyze a three-dimensional point cloud not used for training the deep artificial neural network, and
store the semantically-segmented three-dimensional point cloud on the computer-readable storage medium, wherein the point cloud comprises the set of point labels.
18. The server of claim 16 , wherein one or more of the point labels is determined using a pairwise distance function.
19. The server of claim 16 , wherein one or more of the point labels is determined by selecting an arg max of a vector of labels corresponding to the one or more point labels.
20. The server of claim 15 , wherein the set of point labels includes a type and an elevation.