Learning data generation device, learning data generation method, and program
A learning data generation device for generating learning data for learning a recognizer capable of estimating a contour of a sphere making spinning motion, with high accuracy, the sphere being recorded in a single camera video image, is provided. The learning data generation device includes: a spinning rate estimation unit that receives an input of a learning video image in which motion of a spinning sphere is recorded and an initial value of a size of a contour of the recorded sphere in the video image, sets a plurality of set values of the size of the contour based on the initial value, and obtains an estimated value of a spinning rate of the sphere based on the learning video image, for each of the set values; a contour determination unit that receives an input of a true value of the spinning rate of the sphere, the true value being obtained in advance for the learning video image, and determines at least any of a plurality of the set values respectively corresponding to a plurality of the estimated values selected in order of closeness to the true value, as a determined value of the contour; and a learning data output unit that outputs the learning video image and the determined value as learning data.
1. A learning data generation device comprising:
processing circuitry configured to
receive an input of a learning video image in which motion of a spinning sphere is recorded and an initial value of a size of a contour of the recorded sphere in the video image, set a plurality of set values of the size of the contour based on the initial value, and obtain an estimated value of a spinning rate of the sphere based on the learning video image, for each of the set values;
receive an input of a true value of the spinning rate of the sphere, the true value being obtained in advance for the learning video image, and determine at least one of a plurality of the set values respectively corresponding to a plurality of the estimated values selected in order of closeness to the true value, as a determined value of the contour; and
output the learning video image and the determined value as learning data.
2. The learning data generation device according to claim 1 , wherein
the processing circuitry estimates a spinning state of the sphere by, using the learning video image at a time t and the learning video image at a time t+t c where t c is a predetermined integer of no less than 1, selecting, from among a plurality of hypotheses of the spinning state, a hypothesis of the spinning state, such that a likelihood of an image of the sphere resulting from the sphere in the learning video image at a certain time being spun for t c unit time based on the hypothesis of the spinning state is high.
3. The learning data generation device according to claim 2 , wherein
the processing circuitry estimates the spinning state of the sphere by, using the learning video images at times t 1 , t 2 , . . . , t K and the learning video images at times t 1 +t c , t 2 +t c , . . . , t K +t c , selecting, from among a plurality of hypotheses of the spinning state, a hypothesis of the spinning state, such that a likelihood of an image of the sphere resulting from the sphere in the learning video images at the times t 1 , t 2 , . . . , t K being spun for t c unit time based on the hypothesis of the spinning state is high.
4. The learning data generation device according to claim 2 , wherein
the processing circuitry repeatedly perform processing for, for each of the plurality of hypotheses of the spinning state, calculating a likelihood of an image of the sphere resulting from the sphere in the learning video image at the time t or the learning video images at the times t 1 , t 2 , . . . , t K being spun for t c unit time based on the hypothesis of the spinning state, and processing for newly generating a plurality of likely hypotheses of the spinning state based on the calculated likelihoods.
5. The learning data generation device according to claim 4 , wherein the processing for newly generating a plurality of likely hypotheses of the spinning state based on the calculated likelihoods, the processing being performed by the processing circuitry, is processing for newly generating a plurality of hypotheses by repeating, a plurality of times, processing for determining a hypothesis, from among the plurality of hypotheses of the spinning state, in such a manner that a hypothesis, based on which the calculated likelihood of the hypothesis is higher, is determined with a higher probability and determining a spinning state having a value obtained by addition of a random number to a value of the spinning state in the determined hypothesis, as a new hypothesis.
6. A learning data generation method comprising:
a step of receiving an input of a learning video image in which motion of a spinning sphere is recorded and an initial value of a size of a contour of the recorded sphere in the video image, setting a plurality of set values of the size of the contour based on the initial value, and obtaining an estimated value of a spinning rate of the sphere based on the learning video image, for each of the set values;
a step of receiving an input of a true value of the spinning rate of the sphere, the true value being obtained in advance for the learning video image, and determining at least one of a plurality of the set values respectively corresponding to a plurality of the estimated values selected in order of closeness to the true value, as a determined value of the contour; and
a learning data step of outputting the learning video image and the determined value as learning data.
7. A non-transitory computer readable medium that stores a program that causes a computer to function as the learning data generation device according to claim 1 .
8. A non-transitory computer readable medium that stores a program that causes a computer to function as the learning data generation device according to claim 2 .
9. A non-transitory computer readable medium that stores a program that causes a computer to function as the learning data generation device according to claim 3 .
10. A non-transitory computer readable medium that stores a program that causes a computer to function as the learning data generation device according to claim 4 .
11. A non-transitory computer readable medium that stores a program that causes a computer to function as the learning data generation device according to claim 5 .