US 11,995,150 B2
Information processing method and information processing system
Denis Gudovskiy, Emeryville, CA (US); Alec Hodgkinson, Pescadero, CA (US); Takuya Yamaguchi, Osaka (JP); and Sotaro Tsukizawa, Osaka (JP)
Assigned to PANASONIC INTELLECTUAL PROPERTY CORPORATION OF AMERICA, Torrance, CA (US)
Filed by Panasonic Intellectual Property Corporation of America, Torrance, CA (US)
Filed on Apr. 19, 2021, as Appl. No. 17/234,127.
Application 17/234,127 is a continuation of application No. PCT/JP2019/045997, filed on Nov. 25, 2019.
Claims priority of provisional application 62/818,144, filed on Mar. 14, 2019.
Claims priority of application No. 2019-164054 (JP), filed on Sep. 9, 2019.
Prior Publication US 2021/0241021 A1, Aug. 5, 2021
Int. Cl. G06F 18/214 (2023.01); G06N 5/04 (2023.01); G06N 20/00 (2019.01)
CPC G06F 18/214 (2023.01) [G06N 5/04 (2013.01); G06N 20/00 (2019.01)] 9 Claims
OG exemplary drawing
 
1. An information processing method implemented by a computer, the information processing method comprising:
obtaining first image data and first label information for use in evaluating a neural network that outputs output data including a recognition result indicating a recognition object recognized from input image data, the first image data including a first object whose image has been captured by a first camera, the first label information being information indicating the first object, the neural network being trained by machine learning using pieces of training image data;
obtaining first output data as the output data from the neural network, the first output data including a first recognition result indicating a first recognition object recognized from the first image data by the neural network as a result of inputting the first image data as the input image data into the neural network;
selecting the first image data as falsely recognized image data, when the first recognition object indicated by the first output data obtained does not match the first object indicated by the first label information;
calculating a first contribution representing degrees of contribution of respective pixels to recognition from the first image data by the neural network, the respective pixels being included in the first image data selected as the falsely recognized image data;
obtaining second image data to be a candidate for a new piece of training image data for use in training the neural network, the second image data including a second object whose image has been captured by a second camera;
obtaining second output data as the output data from the neural network, the second output data including a second recognition result indicating a second recognition object recognized from the second image data by the neural network as a result of inputting the second image data as the input image data into the neural network;
calculating a second contribution representing degrees of contribution of respective pixels to recognition from the second image data by the neural network, the respective pixels being included in the second image data;
calculating a degree of similarity between the first contribution and the second contribution;
determining whether to add the second image data to the pieces of training image data, in accordance with the degree of similarity calculated; and
training the neural network by using a new set of pieces of training data obtained by adding, to the pieces of training data, the second data determined to be added.