collecting surface data from a three-dimensional shape as learning data; and

training a set of a prescribed encoder and a prescribed decoder by using the learning data, the training extracting a three-dimensional shape descriptor from the learning data by using the encoder, undoes the three-dimensional shape descriptor by using the decoder, evaluates a difference between a state before using the encoding and a state after using the decoding, and adjusts the encoder and the decoder so as to reduce the error, wherein

the collecting collects, as the surface data, more point cloud data for adjacent mesh faces forming an acuter surface angle for meshes capable of being encoded and decoded by the three-dimensional shape descriptor extractor,

the training trains the encoder and the decoder using the surface data.