Method and apparatus for generating data for estimating three-dimensional (3D) pose of object included in input image, and prediction model for estimating 3D pose of object

1. A method of generating data for estimating a three-dimensional (3D) pose of an object included in an input image using a computer system, the method comprising:

acquiring, using the computer system, the input image including at least one moving object;

estimating, using the computer system, location information about each joint of a plurality of joints of the object included in the input image using a prediction model pretrained to estimate the location information about each of the plurality of joints of the object;

generating, using the computer system, animation data that represents a movement of the object based on the estimated location information; and

generating a virtual character that mimics the movement of the object by mapping the plurality of joints to the virtual character based on the animation data;

wherein the generated virtual character is arranged at a location of an output image output from a user terminal, the location being determined according to a manipulation by a user of the user terminal.

2. The method of claim 1 , wherein

the prediction model is trained using a teacher model pretrained to estimate, from a desired training image, location information about a plurality of joints of an object included in the training image,

the prediction model is a further compressed model compared to the teacher model, and

the prediction model learns the location information about the plurality of joints of the object included in the input image by using the estimated location information about the plurality of joints of the object included in the input image through the teacher model for calculation of a loss function.

3. The method of claim 1 , wherein

the estimating the location information comprises,

generating a heat map and a location map for each joint of the plurality of joints; and

estimating 3D coordinate information about each joint as the location information based on the heat map and the location map,

the heat map represents a likelihood that each joint is present at each pixel location in the heat map, and

the location map includes information that represents a location in an x-axial direction, information that represents a location in a y-axial direction, and information that represents a location in a z-axial direction for each joint.

4. The method of claim 1 , further comprising:

determining erroneously estimated location information in the estimated location information; and

correcting the erroneously estimated location information.

5. The method of claim 3 , further comprising:

determining whether a value on the heat map for each joint corresponding to the 3D coordinate information about each joint is less than a threshold;

determining that the 3D coordinate information is erroneously estimated in response to the value on the heat map corresponding to the 3D coordinate information being less than the threshold; and

correcting the 3D coordinate information that is determined to be erroneously estimated.

6. The method of claim 1 , wherein

the estimated location information includes 3D coordinate information about each of the plurality of joints, and

the animation data includes 3D location information about each joint based on the 3D coordinate information and rotation information about each joint based on 3D coordinate information that is continuously estimated for each joint.

7. The method of claim 6 , further comprising:

determining whether rotation information that represents an abnormal rotation state is present in the rotation information about each joint; and

correcting the rotation information determined to represent the abnormal rotation state.

8. The method of claim 7 , wherein the determining comprises determining that the rotation information represents the abnormal rotation state in response to the rotation information about each joint being out of a range of a preset angle corresponding to a normal movement of the object.

9. The method of claim 1 , further comprising:

calculating global location information that represents a global location of the object based on a focal distance of a camera used to capture the input image and specific location information estimated for at least one of the plurality of joints,

wherein a location movement of the object in the input image is estimated based on the global location information.

10. The method of claim 1 , further comprising:

generating a confidence map associated with a plurality of objects included in the input image, the confidence map including vector information associated with each joint of a plurality of joints of each of the plurality of objects, and

wherein the estimating the location information comprises estimating the location information about each joint of the plurality of joints of each object to be distinguished from that of another object, among the plurality of objects, using the confidence map.

11. The method of claim 1 , wherein

the generating the virtual character comprises generating a plurality of virtual characters that respectively mimic movements of a plurality of objects included in the input image by mapping a plurality of joints of each of the plurality of objects to each of the plurality of virtual characters based on the animation data, and

each of the generated plurality of virtual characters is arranged at a different location of an output image output from a user terminal.

12. The method of claim 1 , wherein

the acquiring comprises a partial image that includes a single object among a plurality of objects in the input image, and

the estimating and the generating are performed in parallel or sequentially with respect to the partial image.

13. The method of claim 1 , wherein

the generating the virtual character comprises generating a plurality of virtual characters that respectively mimic movements of a plurality of objects included in the input image by mapping a plurality of joints of each of the plurality of objects to a corresponding virtual character, and

each of the generated plurality of virtual characters is arranged at a location of the output image determined according to a manipulation by the user of the user terminal.

14. A non-transitory computer-readable record medium storing instructions that, when executed by the computer system, cause the computer system to perform the method of claim 1 .

15. A computer system for generating data for estimating a three-dimensional (3D) pose of an object included in an input image, the computer system comprising:

at least one processor configured to execute computer-readable instruction,

wherein the at least one processor is configured to

acquire the input image including at least one moving object,

estimate location information about each joint of a plurality of joints of the object included in the input image using a prediction model pretrained to estimate the location information about each of the plurality of joints of the object,

generate animation data that represents a movement of the object based on the estimated location information, and

generate a virtual character that mimics the movement of the object by mapping the plurality of joints to the virtual character based on the animation data,

wherein the generated virtual character is arranged at a location of an output image output from a user terminal, the location being determined according to a manipulation by a user of the user terminal.

16. A method of building a prediction model for estimating a three-dimensional (3D) pose of an object included in an input image using a computer system, the method comprising:

training, using the computer system, a teacher model to estimate location information about a plurality of joints of a training object included in a training image;

training, using the computer system, a student model of the teacher model that is a further compressed model compared to the teacher model to learn location information about the plurality of joints of the object included in the training image based on the estimated location information about the plurality of joints of the object included in the training image through the teacher model; and

building, using the computer system, the trained student model as the prediction model for estimating the 3D pose of the object included in the input image.

17. The method of claim 16 , wherein

the training the student model comprises training the student model using a first loss function and a second loss function,

the first loss function is calculated based on

a difference between a first heat map for each joint of the plurality of joints estimated by the student model and a second heat map corresponding to ground truth for each joint, and

a difference between the first heat map and a third heat map for each joint estimated by the teacher model, and

the second loss function is calculated based on

a difference between a first location map for each joint estimated by the student model and a second location map corresponding to ground truth for each joint, and

a difference between the first location map and a third location map for each joint estimated by the teacher model.

18. The method of claim 17 , wherein

the difference between the first heat map and the second heat map and the difference between the first heat map and the third heat map are included in the first loss function with a first mixture ratio, and

the difference between the first location map and the second location map and the difference between the first location map and the third location map are included in the second loss function with a second mixture ratio.