Training system and analysis system

1. A training system, comprising:

processing circuitry configured to

use a human body model to generate a first image, the human body model modeling a human body and being three-dimensional and virtual;

generate a teacher image by annotating body parts of the human body model in the first image;

use the first image to generate a second image including noise; and

use the second image and the teacher image to train a first model including a neural network,

the neural network including

a first block reducing an input image and outputting a feature map,

a first residual block receiving input of the feature map output from the first block,

a second residual block receiving input of a feature map output from the first residual block, and outputting a feature map having a same size as the input feature map and the feature map output from the first residual block,

a second block receiving input of the feature map output from the first residual block and the feature map output from the second residual block,

a third block receiving input of the feature map output from the first block and a feature map output from the second block, and

a fourth block enlarging a feature map output from the third block and outputting an output image.

2. The training system according to claim 1 , wherein the first image includes depth information.

3. The training system according to claim 1 , wherein

the processing circuitry:

sets positions of the body parts of the human body model to match position data of positions of body parts of a person in real space; and

generates the first image by using the human body model in which the positions of the body parts are set.

4. The training system according to claim 1 , wherein the first image includes the human body model imaged from above.

5. The training system according to claim 1 , wherein the processing circuitry generates the first image by also using an environment model modeling an environment in real space.

6. An analysis system, comprising processing circuitry that performs inputting the input image of a first person in real space to the first model trained by the training system according to claim 1 to acquire the output image in which body parts of the first person are annotated.

7. The analysis system according to claim 6 ,

the processing circuitry

acquiring a plurality of the output images by inputting, to the first model, a plurality of the input images showing a movement of the first person in a first work, and

using a plurality of skeleton information based on the plurality of output images to generate time-series data of a movement of the body parts over time, and calculating a period of a movement of the first person based on the time-series data.

8. A training method, comprising:

using a human body model to generate a first image, the human body model modeling a human body and being three-dimensional and virtual;

generating a teacher image by annotating body parts of the human body model in the first image;

using the first image to generate a second image including noise; and

using the second image and the teacher image to train a first model including a neural network,

the neural network including

a first block reducing an input image and outputting a feature map,

a first residual block receiving input of the feature map output from the first block,

a second residual block receiving input of a feature map output from the first residual block, and outputting a feature map having a same size as the input feature map and the feature map output from the first residual block,

a second block receiving input of the feature map output from the first residual block and the feature map output from the second residual block,

a third block receiving input of the feature map output from the first block and a feature map output from the second block, and

a fourth block enlarging a feature map output from the third block and outputting an output image.

9. The training method according to claim 8 , wherein the first image includes depth information.

10. The training method according to claim 8 , wherein

in generating the teacher image,

positions of the body parts of the human body model are set to match position data of positions of body parts of a person in real space; and

the first image is generated by using the human body model in which the positions of the body parts are set.

11. The training method according to claim 8 , wherein the first image includes the human body model imaged from above.

12. The training method according to claim 8 , wherein the first image is generated by also using an environment model modeling an environment in real space.

13. An analysis method, comprising inputting the input image of a first person in real space to the first model trained by the training method according to claim 8 to acquire the output image in which body parts of the first person are annotated.

14. The analysis method according to claim 13 , further comprising:

acquiring a plurality of the output images by inputting, to the first model, a plurality of the input images showing a movement of the first person in a first work;

using a plurality of skeleton information based on the plurality of output images to generate time-series data of a movement of the body parts over time; and

calculating a period of a movement of the first person based on the time-series data.

15. A non-transitory computer-readable storage medium storing a program, the program causing a computer to perform:

using a human body model to generate a first image, the human body model modeling a human body and being three-dimensional and virtual;

generating a teacher image by annotating body parts of the human body model in the first image;

using the first image to generate a second image including noise; and

using the second image and the teacher image to train a first model including a neural network,

the neural network including

a first block reducing an input image and outputting a feature map,

a first residual block receiving input of the feature map output from the first block,

a second residual block receiving input of a feature map output from the first residual block, and outputting a feature map having a same size as the input feature map and the feature map output from the first residual block,

a second block receiving input of the feature map output from the first residual block and the feature map output from the second residual block,

a third block receiving input of the feature map output from the first block and a feature map output from the second block, and

a fourth block enlarging a feature map output from the third block and outputting an output image.

16. The non-transitory computer-readable storage medium according to claim 15 , wherein the first image includes depth information.

17. The non-transitory computer-readable storage medium according to claim 15 , wherein

in generating the teacher image, the program causes the computer to perform,

setting positions of the body parts of the human body model to match position data of positions of body parts of a person in real space, and

generating the first image by using the human body model in which the positions of the body parts are set.

18. The non-transitory computer-readable storage medium according to claim 15 , wherein the first image includes the human body model imaged from above.

19. The non-transitory computer-readable storage medium according to claim 15 , wherein the first image is generated by also using an environment model modeling an environment in real space.

20. The non-transitory computer-readable storage medium according to claim 15 , wherein

the program further causes the computer to perform

inputting the input image of a first person in real space to the trained first model to acquire the output image in which body parts of the first person are annotated,

acquiring a plurality of the output images by inputting, to the first model, a plurality of the input images showing a movement of the first person in a first work,

using a plurality of skeleton information based on the plurality of output images to generate time-series data of a movement of the body parts over time, and

calculating a period of a movement of the first person based on the time-series data.