Two-stage body pose estimation
In one implementation, a method of body pose estimation is performed at a device including one or more processors and non-transitory memory. The method includes obtaining a plurality of two-dimensional images of a body in a three-dimensional environment at a respective plurality of times. The method includes determining, for each of the plurality of two-dimensional images, the two-dimensional location in the two-dimensional image of one or more joints of the body at the respective plurality of times. The method includes determining, based on the two-dimensional locations, a plurality of three-dimensional locations in the three-dimensional environment of the one or more joints of the body at the respective plurality of times. The method includes determining, based on the three-dimensional locations, a plurality of updated three-dimensional locations in the three-dimensional environment of the one or more joints of the body at the respective plurality of times.
1. A method comprising:
at a device including one or more processors and non-transitory memory:
capturing, via a camera, a plurality of two-dimensional images of a body in a three-dimensional environment at a respective plurality of times;
determining, for each of the plurality of two-dimensional images, two-dimensional locations in the two-dimensional image of one or more joints of the body at the respective plurality of times;
determining, based on the two-dimensional locations, a plurality of three-dimensional location values in the three-dimensional environment of the one or more joints of the body at the respective plurality of times; and
modifying the plurality of three-dimensional location values to generate a plurality of updated three-dimensional location values in the three-dimensional environment of the one or more joints of the body at the respective plurality of times.
2. The method of claim 1 , wherein determining the plurality of three-dimensional location values includes applying a first neural network to the two-dimensional locations.
3. The method of claim 2 , wherein the first neural network is trained on data including joint occlusions.
4. The method of claim 2 , wherein determining the plurality of updated three-dimensional location values includes applying a second neural network to the three-dimensional location values.
5. The method of claim 4 , wherein determining the plurality of updated three-dimensional location values includes applying the second neural network to the three-dimensional location values iteratively a number of times.
6. The method of claim 5 , wherein the number of times is based on a computational budget.
7. The method of claim 4 , wherein the second neural network is trained on a loss function including a position term and a velocity term.
8. The method of claim 1 , wherein modifying the plurality of three-dimensional location values comprises:
temporally refining the plurality of three-dimensional location values so that the plurality of updated three-dimensional location values has a greater temporal consistency than the plurality of three-dimensional location values.
9. A device comprising:
a non-transitory memory; and
one or more processors to:
capture, via a camera, a plurality of two-dimensional images of a body in a three-dimensional environment at a respective plurality of times;
determine, for each of the plurality of two-dimensional images, two-dimensional locations in the two-dimensional image of one or more joints of the body at the respective plurality of times;
determine, based on the two-dimensional locations, a plurality of three-dimensional location values in the three-dimensional environment of the one or more joints of the body at the respective plurality of times; and
modify the three-dimensional location values to generate a plurality of updated three-dimensional location values in the three-dimensional environment of the one or more joints of the body at the respective plurality of times.
10. The device of claim 9 , wherein the one or more processors are to determine the plurality of three-dimensional location values by applying a first neural network to the two-dimensional locations.
11. The device of claim 10 , wherein the first neural network is trained on data including joint occlusions.
12. The device of claim 10 , wherein the one or more processors are to determine the plurality of updated three-dimensional location values by applying a second neural network to the three-dimensional location values.
13. The device of claim 12 , wherein the one or more processors are to determine the plurality of updated three-dimensional location values by applying the second neural network to the three-dimensional location values iteratively a number of times.
14. The device of claim 13 , wherein the number of times is based on a computational budget.
15. The device of claim 12 , wherein the second neural network is trained on a loss function including a position term and a velocity term.
16. The device of claim 9 , wherein modifying the plurality of three-dimensional location values comprises:
temporally refining the plurality of three-dimensional location values so that the plurality of updated three-dimensional location values is less jittery than the plurality of three-dimensional location values.
17. A non-transitory computer-readable medium having instructions encoded thereon, which when executed by one or more processors of a device, cause the device to:
capture, via a camera, a plurality of two-dimensional images of a body in a three-dimensional environment at a respective plurality of times;
determine, for each of the plurality of two-dimensional images, two-dimensional locations in the two-dimensional image of one or more joints of the body at the respective plurality of times;
determine, based on the two-dimensional locations, a plurality of three-dimensional location values in the three-dimensional environment of the one or more joints of the body at the respective plurality of times; and
modify the plurality of three-dimensional location values to generate a plurality of updated three-dimensional location values in the three-dimensional environment of the one or more joints of the body at the respective plurality of times.
18. The non-transitory computer-readable medium of claim 17 , wherein instructions cause the device to determine the plurality of three-dimensional location values by applying a first neural network to the two-dimensional locations.
19. The non-transitory computer-readable medium of claim 18 , wherein the first neural network is trained on data including joint occlusions.
20. The non-transitory computer-readable medium of claim 18 , wherein the instructions cause the device to determine the plurality of updated three-dimensional location values by applying a second neural network to the three-dimensional location values.
21. The non-transitory computer-readable medium of claim 20 , wherein the instructions cause the device to determine the plurality of updated three-dimensional location values by applying the second neural network to the three-dimensional location values iteratively a number of times.
22. The non-transitory computer-readable medium of claim 20 , wherein the second neural network is trained on a loss function including a position term and a velocity term.
23. The non-transitory computer-readable medium of claim 17 , wherein modifying the plurality of three-dimensional location values comprises:
utilizing a temporal refiner to temporally refine the plurality of three-dimensional location values so that the plurality of updated three-dimensional location values has less jitter and greater temporal consistency than the plurality of three-dimensional location values.