METHOD FOR GENERATING THREE-DIMENSIONAL POINT CLOUDS Russian patent published in 2021 - IPC G06N3/08 

FIELD: neural networks.

SUBSTANCE: invention relates to a method for training a neural network model to generate three-dimensional point clouds. In the method, a point cloud is generated by combining several autoencoders and several generative adversarial networks, while a point cloud is fed to the input of each autoencoder, which is encoded into a vector in a hidden space, generative adversarial networks are trained on a vector of random values ​​in a hidden space, where the input the discriminator of the generative adversarial network serves the generated vector in the hidden space from the generator of the generative adversarial network, as well as the vector in the hidden space for the real object from the encoder of the autoencoder; the generated vector is fed to the input of each decoder, which is generated in the hidden space from the generator of the generative adversarial network after training the combined autoencoder and the generative adversarial network, which is decoded into a matrix, at the output of each decoder a point cloud is obtained; in this case, all autoencoders and all generative adversarial networks are combined into a hierarchical structure, repeating the execution of the training stages of the combined autoencoder and the generative adversarial network, each time after which the generated vector in the hidden space from the generator of the generative adversarial network is doubled by means of the operator of doubling the number of points, and is transmitted for training to the next neural network model of the combined autoencoder and generative-adversarial network, the output of which is a denser and higher-quality point cloud.

EFFECT: technical result consists in generating higher quality three-dimensional point clouds.

1 cl, 7 dwg