Hand pose estimation for machine learning based gesture recognition
The technology disclosed performs hand pose estimation on a so-called “joint-by-joint” basis. So, when a plurality of estimates for the 28 hand joints are received from a plurality of expert networks (and from master experts in some high-confidence scenarios), the estimates are analyzed at a joint level and a final location for each joint is calculated based on the plurality of estimates for a particular joint. This is a novel solution discovered by the technology disclosed because nothing in the field of art determines hand pose estimates at such granularity and precision. Regarding granularity and precision, because hand pose estimates are computed on a joint-by-joint basis, this allows the technology disclosed to detect in real time even the minutest and most subtle hand movements, such a bend/yaw/tilt/roll of a segment of a finger or a tilt an occluded finger, as demonstrated supra in the Experimental Results section of this application.
1. A method of preparing sample hand positions for training of neural network systems, the method including:
accessing simulation parameters that specify at least one of:
a range of hand positions and position sequences,
a range of hand anatomies, including palm size, fattiness, stubbiness, and skin tone, and
a range of backgrounds;
accessing a camera perspective specification that specifies one or more of:
a focal length,
a field of view of the camera,
a wavelength sensitivity, and
artificial lighting conditions;
generating a plurality of hand position-hand anatomy-background simulations, each simulation labeled with hand position parameters, including joint locations of joints of the hand in three dimensions (3D), in a ground truth vector for training a convolutional neural network, the simulations organized in sequences;
applying the camera perspective specification to render from the simulations at least a corresponding set of simulated hand position images; and
saving the simulated hand position images with one or more labelled hand position parameters from the corresponding simulations.
2. The method of claim 1 , wherein the simulated hand position images are stereoscopic images with depth map information.
3. The method of claim 1 , wherein the simulated hand position images are binocular pairs of images.
4. The method of claim 1 , wherein the hand position parameters are a plurality of joint locations in three-dimensional (3D) space.
5. The method of claim 1 , wherein the hand position parameters are a plurality of joint angles in three-dimensional (3D) space.
6. The method of claim 1 , wherein the hand position parameters are a plurality of hand skeleton segments in three-dimensional (3D) space.
7. The method of claim 1 , the method further including:
extracting stereoscopic hand boundaries for the simulated hand position images and aligning the hand boundaries with hand centers included in the simulated hand position images;
generating translated, rotated and scaled variants of the hand boundaries and applying Gaussian jittering to the variants;
extracting hand regions from one or more jittered variants of the hand boundaries;
computing ground truth pose vectors for the hand regions; and
storing the pose vectors in tangible machine readable memory as output labels for the simulated hand position images.
8. The method of claim 7 , further including generating three-dimensional (3D) simulated hands in mesh models and/or capsule hand skeleton models.
9. The method of claim 7 , further including generating a simulated coordinate system to determine hand position parameters of a simulated hand in three-dimensional (3D).
10. The method of claim 7 , further including generating a simulated perspective of a simulated gesture recognition system to determine hand position parameters of a simulated hand in three-dimensional (3D).
11. The method of claim 7 , wherein the ground truth pose vectors are 84 dimensional representing 28 hand joints in three-dimensional (3D) space.
12. A non-transitory computer readable storage medium impressed with computer program instructions to prepare sample hand positions for training of neural network systems, the instructions, when executed on a processor, implement a method comprising:
accessing simulation parameters that specify at least one of:
a range of hand positions and position sequences,
a range of hand anatomies, including palm size, fattiness, stubbiness, and skin tone, and
a range of backgrounds;
accessing a camera perspective specification that specifies one or more of:
a focal length,
a field of view of the camera,
a wavelength sensitivity, and
artificial lighting conditions;
generating a plurality of hand position-hand anatomy-background simulations, each simulation labeled with hand position parameters, including joint locations of joints of the hand in three dimensions (3D), in a ground truth vector for training a convolutional neural network, the simulations organized in sequences;
applying the camera perspective specification to render from the simulations at least a corresponding set of simulated hand position images; and
saving the simulated hand position images with one or more labelled hand position parameters from the corresponding simulations.
13. The non-transitory computer readable storage medium of claim 12 , wherein the simulated hand position images are stereoscopic images with depth map information.
14. The non-transitory computer readable storage medium of claim 12 , wherein the simulated hand position images are binocular pairs of images.
15. The non-transitory computer readable storage medium of claim 12 , wherein the hand position parameters include parameters selected from one or more of: (i) a plurality of joint locations in three-dimensional (3D) space; (ii) a plurality of joint angles in three-dimensional (3D) space; and (iii) a plurality of hand skeleton segments in three-dimensional (3D) space.
16. The non-transitory computer readable storage medium of claim 12 , further including instructions which when executed by the processor, implement:
extracting stereoscopic hand boundaries for the simulated hand position images and aligning the hand boundaries with hand centers included in the simulated hand position images;
generating translated, rotated and scaled variants of the hand boundaries and applying Gaussian jittering to the variants;
extracting hand regions from one or more jittered variants of the hand boundaries;
computing ground truth pose vectors for the hand regions; and
storing the pose vectors in tangible machine readable memory as output labels for the simulated hand position images.
17. The non-transitory computer readable storage medium of claim 16 , further including generating three-dimensional (3D) simulated hands in mesh models and/or capsule hand skeleton models.
18. The non-transitory computer readable storage medium of claim 16 , further including generating a simulated coordinate system to determine hand position parameters of a simulated hand in three-dimensional (3D).
19. The non-transitory computer readable storage medium of claim 16 , further including generating a simulated perspective of a simulated gesture recognition system to determine hand position parameters of a simulated hand in three-dimensional (3D).
20. The non-transitory computer readable storage medium of claim 16 , wherein the ground truth pose vectors are 84 dimensional representing 28 hand joints in three-dimensional (3D) space.
21. A system including one or more processors coupled to memory, the memory loaded with computer instructions, which instructions, when executed on the one or more processors, implement actions comprising the method of claim 1 :
accessing simulation parameters that specify at least one of:
a range of hand positions and position sequences,
a range of hand anatomies, including palm size, fattiness, stubbiness, and skin tone, and
a range of backgrounds;
accessing a camera perspective specification that specifies one or more of:
a focal length,
a field of view of the camera,
a wavelength sensitivity, and
artificial lighting conditions;
generating a plurality of hand position-hand anatomy-background simulations, each simulation labeled with hand position parameters, including joint locations of joints of the hand in three dimensions (3D), in a ground truth vector for training a convolutional neural network, the simulations organized in sequences;
applying the camera perspective specification to render from the simulations at least a corresponding set of simulated hand position images; and
saving the simulated hand position images with one or more labelled hand position parameters from the corresponding simulations for use in training a hand position recognition system.
22. The method of claim 1 , wherein the plurality of hand position-hand anatomy-background simulations includes simulations representing at least two different hands with different respective hand anatomies.