Learning device, learning method, and recording medium
A data acquisition means acquires source domain data and target domain data. An alignment means performs an alignment which converts the source domain data and the target domain data into images of a predetermined reference angle. A feature extraction means extracts local features of the source domain data and the target domain data. A classification means classifies a class based on the local features of the source domain data and the target domain data after the alignment. A learning means trains the feature extraction means based on the local features of the source domain data and the target domain data after the alignment and a classification result of the class.
1 . A learning device comprising:
a memory storing instructions; and
one or more processors configured to execute the instructions to:
acquire source domain data and target domain data;
perform an alignment which converts the source domain data and the target domain data into images of a predetermined reference angle;
extract local features of the source domain data and the target domain data by using a feature extraction;
classify a class based on the local features of the source domain data and the target domain data after the alignment; and
train the feature extraction based on the local features of the source domain data and the target domain data after the alignment and a classification result of the class.
2 . The learning device according to claim 1 , wherein the processor is further configured to
estimate conversion parameters for converting input image data into image data of the reference angle; and
convert the input image data by using the conversion parameters.
3 . The learning device according to claim 2 , wherein the conversion parameter estimation has been trained by using an image group generated by rendering source domain data of a specific class under a different condition and a reference image generated by rendering the source domain data of the specific class at the reference angle.
4 . The learning device according to claim 3 , wherein the conversion parameter estimation has been trained to minimize an error between the image group converted by using the conversion parameters and the reference image.
5 . The learning device according to claim 1 , wherein the processor is further configured to
estimate conversion parameters for converting input image features into image features derived from an image of the reference angle by a conversion parameter estimation; and
convert the input image features by using the conversion parameters.
6 . The learning device according to claim 5 , wherein the conversion parameter estimation has been trained by using image features derived from an image group generated by rendering source domain data of a specific class under a different condition and image features derived from a reference image generated by rendering the source domain data of the specific class at the reference angle.
7 . The learning device according to claim 6 , wherein the conversion parameter estimation is trained to minimize an error between image features derived from the image group converted by using the conversion parameters and image features derived from the reference image.
8 . The learning device according to claim 1 , wherein the rendering to generate the image group includes at least one of a change of an angle of a line of sight with respect to an object in an image, a change of a distance of the object in a depth direction in the image, a parallel movement of the object, an addition or a change of color of the object, an addition or a change of a pattern of the object, an addition of illumination with respect to the object, an addition of a background of the object, and an addition of noise.
9 . The learning device according to claim 1 , wherein the feature extraction extracts a plurality of image features from input image data, and outputs respective correlations among the plurality of image features as the local features.
10 . The learning device according to claim 1 , wherein the processor minimizes an error between a classification result of the class and a correct answer label, and trains the feature extraction so that the local features of the source domain data and the local features of the target domain data are closer for the same class, and the local features extracted from the target and the local features extracted from source domain data are farther apart for different classes.
11 . A learning method, comprising:
acquiring source domain data and target domain data;
performing an alignment which converts the source domain data and the target domain data into images of a predetermined reference angle;
extracting local features of the source domain data and the target domain data by a feature extraction;
classifying a class based on the local features of the source domain data and the target domain data after the alignment; and
training the feature extraction based on the local features of the source domain data and the target domain data after the alignment and a classification result of the class.
12 . A non-transitory computer-readable recording medium storing a program, the program causing a computer to perform a process comprising:
acquiring source domain data and target domain data;
performing an alignment which converts the source domain data and the target domain data into images of a predetermined reference angle;
extracting local features of the source domain data and the target domain data by a feature extraction;
classifying a class based on the local features of the source domain data and the target domain data after the alignment; and
training the feature extraction based on the local features of the source domain data and the target domain data after the alignment and a classification result of the class.