THREE-DIMENSIONAL DATA EVALUATION MODEL CONSTRUCTION AND EVALUATION METHOD, THREE-DIMENSIONAL DATA EVALUATION MODEL CONSTRUCTION DEVICE, THREE-DIMENSIONAL DATA EVALUATION DEVICE, AND COMPUTER PROGRAM

摘要

To perform semi-supervised learning on three-dimensional data in which the number of data available as a learning data set is limited even though the amount of data is large, while preventing overlearning.SOLUTION: Provided is a method including at least a part of the following steps of: learning a denoising function by passing three-dimensional data through a convolutional ladder network; extracting residual data between reconstructed data based on a learning result and original three-dimensional data; constructing a plurality of evaluation nodes having different viewpoints by using a neural network treating the residual data as an input, thereby calibrating each evaluation node by using VAT; inputting residual data for new three-dimensional data to the plurality of evaluation nodes to obtain outputs; and determining the reliability of the outputs according to consistency between the plurality of outputs. Also, provided are a device for executing the method, and a program for configuring the device.SELECTED DRAWING: Figure 1