Non-transitory computer-readable storage medium for storing position detection program, position detection method, and position detection apparatus

1. A non-transitory computer-readable storage medium storing a position detection program which causes a processor to perform processing for object recognition, the processing comprising:

acquiring, as training data, a plurality of pieces of three-dimensional shape data of simple shapes that are not similar to each other, each of the plurality of pieces of three-dimensional shape data being three-dimensional shape data representing a simple shape obtained from a three-dimensional model, the simple shape being a shape simplified more than a shape of a target object;

carrying out learning by using, as an input to a neural network, the simple shape represented by each of the plurality of acquired pieces of three-dimensional shape data;

acquiring an image obtained by imaging by an imaging device; and

detecting a position of the target object from the acquired image by using a first learning model generated based on the learning.

2. The position detection program according to claim 1 , the processing further comprising:

detecting a gripping position of the object from the position.

3. The position detection program according to claim 1 , wherein

the first learning model is an ensemble learning model based on a plurality of second learning models generated through carrying out learning by using each of the plurality of pieces of data.

4. The position detection program according to claim 3 , wherein

the second learning model is a convolutional neutral network using deep learning and creates, from the image, a plurality of bounding boxes that represent a detected position of an object as position candidates in such a manner as to associate the bounding boxes with reliability that represents likelihood of a detected object, and

the ensemble learning model acquires a second number of position candidates from each of a first number of second learning models and, when an overlapping region exists in bounding boxes of position candidates in a number obtained by multiplying the first number by the second number, the ensemble learning model groups position candidates having the overlapping region and employs sum of reliabilities of the grouped position candidates as reliability of the group to output a bounding box of the group or position candidate with highest reliability as information on the position of the object.

5. The position detection program according to claim 1 , wherein

the pieces of data are pieces of data created from three-dimensional models of the simple shapes by a simulator.

6. The position detection program according to claim 1 , wherein

the simple shapes include a circular column, a rectangular parallelepiped, a cube, a sphere, and a spring.

7. A position detection method comprising:

by a computer,

acquiring, as training data, a plurality of pieces of three-dimensional shape data of simple shapes that are not similar to each other, each of the plurality of pieces of three-dimensional shape data being three-dimensional shape data representing a simple shape obtained from a three-dimensional model, the simple shape being a shape simplified more than a shape of a target object;

carrying out learning by using, as an input to a neural network, the simple shape represented by each of the plurality of acquired pieces of three-dimensional shape data;

acquiring an image obtained by imaging by an imaging device; and

detecting a position of the target object from the acquired image by using a first learning model generated based on the learning.

8. A position detection apparatus comprising:

memory; and

processor circuitry coupled to the memory, the processor circuitry being configured to:

acquire, as training data, a plurality of pieces of three-dimensional shape data of simple shapes that are not similar to each other, each of the plurality of pieces of three-dimensional shape data being three-dimensional shape data representing a simple shape obtained from a three-dimensional model, the simple shape being a shape simplified more than a shape of a target object;

carry out learning by using, as an input to a neural network, the simple shape represented by each of the plurality of acquired pieces of three-dimensional shape data;

acquire an image obtained by imaging by an imaging device; and

detect a position of the target object from the acquired image by using a first learning model generated based on the learning.