Machine learning systems and methods of estimating body shape from images
Disclosed is a method including receiving an input image including a human, predicting, based on a convolutional neural network that is trained using examples consisting of pairs of sensor data, a corresponding body shape of the human and utilizing the corresponding body shape predicted from the convolutional neural network as input to another convolutional neural network to predict additional body shape metrics.
1. A computer implemented method comprising:
receiving an image depicting a human;
generating a plurality of first training data pairs, wherein each first training data pair includes (i) a synthetic image depicting a computer-generated body model positioned in front of an image background, wherein the computer-generated body model has a texture map applied thereto, and (ii) a synthetic ground truth image representing a shape of the computer-generated body model depicted in the synthetic image;
training a first convolutional neural network (“CNN”) to associate the synthetic ground truth image with the synthetic image included in the plurality of the first training data pairs;
predicting a ground truth image from the image depicting the human using the trained first CNN, wherein the ground truth image represents a body shape of the human;
predicting a body metric for the body shape of the human from the ground truth image using a second CNN; and
generating content for output via a user interface, wherein the content is generated based at least partly on the body metric predicted using the second CNN.
2. The computer implemented method of claim 1 , wherein the body metric comprises a clothing size, and wherein the content generated for output via the user interface comprises a recommendation of a clothing item available in the clothing size.
3. The computer implemented method of claim 1 , wherein the body metric comprises a health condition, the wherein the content generated for output via the user interface comprises information regarding the health condition.
4. The computer implemented method of claim 1 , wherein the first CNN is additionally trained based on one or more real images, wherein each of the real images depicts a real human.
5. The computer implemented method of claim 1 , wherein the computer-generated body model comprises a three-dimensional (“3D”) body model, the method further comprising, for each first training data pair of the plurality of first training data pairs, applying a texture map to the 3D body model.
6. The computer implemented method of claim 1 , further comprising:
accessing a plurality of second training data pairs, wherein each second training data pair includes (i) the synthetic ground truth image of a different one of the first training data pairs and (ii) the body metric corresponding to the synthetic ground truth image; and
training the second CNN to associate the body metric with the synthetic ground truth image included in the plurality of the second training data pairs.
7. A non-transitory computer-readable medium including instructions that, when executed by one or more processors, cause the one or more processors to at least:
receive an image depicting a human;
generate a plurality of first training data pairs, wherein each first training data pair includes (i) a synthetic image depicting a computer-generated body model positioned in front of an image background, wherein the computer-generated body model has a texture map applied thereto, and (ii) a synthetic ground truth image representing a shape of the computer-generated body model depicted in the synthetic image;
train a first convolutional neural network (“CNN”) to associate the synthetic ground truth image with the synthetic image included in the plurality of the first training data pairs;
predict a ground truth image from the image depicting the human using the trained first CNN, wherein the ground truth image represents a body shape of the human;
predict a body metric for the body shape of the human from the ground truth image using a second CNN; and
generate content for output via a user interface, wherein the content is generated based at least partly on the body metric predicted using the second CNN.
8. The non-transitory computer-readable medium of claim 7 , wherein the body metric comprises a clothing size, and wherein the non-transitory computer-medium further includes instructions for generating the user interface to present a recommendation of a clothing item available in the clothing size.
9. The non-transitory computer-readable medium of claim 8 , wherein the body metric comprises a health condition, and wherein the non-transitory computer-medium further includes instructions for generating the user interface to present information regarding the health condition.
10. The non-transitory computer-readable medium of claim 8 , wherein the first CNN is additionally trained based on one or more real images, wherein each of the real images depicts a real human.
11. The non-transitory computer-readable medium of claim 7 , wherein the computer-generated body model comprises a three-dimensional (“3D”) body model, and wherein the non-transitory computer-medium further includes instructions for applying a texture map to the 3D body model for each first training data pair of the plurality of first training data pairs.
12. The non-transitory computer-readable medium of claim 7 further including instructions that, when executed by the one or more processors, further cause the one or more processors to at least:
access a plurality of second training data pairs, wherein each second training data pair includes (i) the synthetic ground truth image of a different one of the first training data pairs, and (ii) the body metric corresponding to the synthetic ground truth image; and
train the second CNN to associate the body metric with the synthetic ground truth image included in the plurality of the second training data pairs.
13. A system comprising:
a memory storing executable instructions; and
one or more processors in communication with the memory, the one or more processors programmed by the executable instructions to at least:
generate a plurality of first training data pairs, wherein each first training data pair includes (i) a synthetic image depicting a three-dimensional (“3D”) body model and (ii) a synthetic ground truth image representing a shape of the 3D body model depicted in the synthetic image;
train a first convolutional neural network (“CNN”) to associate the synthetic ground truth image with the synthetic image included in the plurality of the first training data pairs;
predict a ground truth image from an image depicting a human using the trained first CNN, wherein the ground truth image represents a body shape of the human;
predict a body metric for the body shape of the human from the ground truth image using a second CNN; and
generate content for output via a user interface, wherein the content is generated based at least partly on the body metric predicted using the second CNN.
14. The system of claim 13 , wherein the 3D body model has a texture map applied thereto.
15. The system of claim 13 , wherein the body metric comprises clothing size, and wherein the one or more processors are further programmed to at least generate the user interface to present a recommendation of a clothing item available in the clothing size.
16. The system of claim 13 , wherein the body metric comprises a health condition, and wherein the one or more processors are further programmed to at least generate the user interface to present information regarding the health condition.
17. The system of claim 16 , wherein the one or more processors are further programmed to at least:
determine that the first CNN is not performing sufficiently;
access a plurality of training data pairs, wherein each training data pair includes (i) a real image depicting a real human, and (ii) a real ground truth image corresponding to the real image and representing a body shape of the real human; and
with each training data pair of the plurality of training data pairs, refine the first CNN by further training the first CNN to predict the real ground truth image from the real image.
18. The system of claim 13 , wherein the synthetic ground truth image includes at least one of: (i) a segmented image identifying pixels corresponding to specific body parts of the 3D body model, or (ii) a depth map representing at least one distance between a surface of the 3D body model and a virtual camera position from which the synthetic image is generated.