Using iterative 3D-model fitting for domain adaptation of a hand-pose-estimation neural network
Described is a solution for an unlabeled target domain dataset challenge using a domain adaptation technique to train a neural network using an iterative 3D model fitting algorithm to generate refined target domain labels. The neural network supports the convergence of the 3D model fitting algorithm and the 3D model fitting algorithm provides refined labels that are used for training of the neural network. During real-time inference, only the trained neural network is required. A convolutional neural network (CNN) is trained using labeled synthetic frames (source domain) with unlabeled real depth frames (target domain). The CNN initializes an offline iterative 3D model fitting algorithm capable of accurately labeling the hand pose in real depth frames. The labeled real depth frames are used to continue training the CNN thereby improving accuracy beyond that achievable by using only unlabeled real depth frames for domain adaptation.
1. A method comprising:
training a first neural network using samples from a source domain;
implementing domain adaptation of the first neural network from the source domain to a target domain where labels are not available, comprising a feedback loop whereby:
a) the first neural network infers labels for target domain samples;
b) the labels for the target domain samples are refined using a generative iterative model fitting process to produce refined labels for the target domain; and
c) the refined labels for the target domain are used for training of the first neural network using backpropagation of errors; and
using a second neural network for inverse kinematics that is used as part of the generative iterative model fitting process.
2. The method as in claim 1 , wherein the generative iterative model fitting process attempts to minimize error between pixels in a synthetic frame and a real frame as the synthetic frames are generated using a computer graphics rendering technique.
3. The method as in claim 1 , wherein the generative iterative model fitting process uses a genetic algorithm.
4. The method as in claim 1 , wherein the source domain samples are generated using a computer graphics rendering technique.
5. The method as in claim 1 , wherein the target domain samples are generated using a camera.
6. The method as in claim 1 , further comprising:
optimizing the first neural network to minimize key-point error from frames for which key-point labels are known.
7. The method as in claim 6 , further comprising:
cropping and resampling frames so that they are centered and have a normalized scale in pixels.
8. The method as in claim 6 , wherein the first neural network is also trained using unlabeled target domain samples, and wherein a domain discriminator neural network and an adversarial loss is used to learn domain invariant features.
9. The method as in claim 3 , further comprising:
using the first neural network with injected noise to generate a pose ensemble for each real depth frame.
10. The method as in claim 9 , further comprising:
using the pose ensemble to initialize a genetic algorithm.
11. The method as in claim 1 , further comprising:
using a forward kinematics model and random pose generator to generate a labeled dataset that is used for training of the second neural network.
12. The method as in claim 1 , further comprising:
using a separate dense neural network for each finger.
13. The method as in claim 4 , further comprising:
using a random pose generator to set a pose of a 3D model for each sample in a synthetic source domain dataset.
14. The method as in claim 12 , further comprising:
adding Gaussian noise to pose key-point inputs while training the second neural network.
15. The method as in claim 8 , wherein the domain discriminator neural network uses global average pooling so that only small-scale domain-invariant features are learned.
16. A method comprising:
training a first neural network using samples from a source domain;
implementing domain adaptation of the first neural network from the source domain to a target domain where labels are not available, comprising a feedback loop whereby:
a) the first neural network infers labels for target domain samples;
b) the labels for the target domain samples are refined using a generative iterative model fitting process to produce refined labels for the target domain; and
c) the refined labels for the target domain are used for training of the first neural network using backpropagation of errors; and
using a second neural network that is used as part of the generative iterative model fitting process.
17. The method as in claim 16 , wherein the generative iterative model fitting process attempts to minimize error between pixels in a synthetic frame and a real frame as the synthetic frames are generated using a computer graphics rendering technique.
18. The method as in claim 16 , wherein the generative iterative model fitting process uses a genetic algorithm.
19. The method as in claim 16 , wherein the source domain samples are generated using a computer graphics rendering technique.
20. The method as in claim 16 , wherein the target domain samples are generated using a camera.
21. The method as in claim 16 , further comprising:
optimizing the first neural network to minimize key-point error from frames for which key-point labels are known.
22. The method as in claim 21 , further comprising:
cropping and resampling frames so that they are centered and have a normalized scale in pixels.
23. The method as in claim 21 , wherein the first neural network is also trained using unlabeled target domain samples, and wherein a domain discriminator neural network and an adversarial loss is used to learn domain invariant features.
24. The method as in claim 18 , further comprising:
using the first neural network with injected noise to generate a pose ensemble for each real depth frame.
25. The method as in claim 24 , further comprising:
using the pose ensemble to initialize a genetic algorithm.
26. The method as in claim 16 , further comprising:
using a forward kinematics model and random pose generator to generate a labeled dataset that is used for training of the second neural network.
27. The method as in claim 16 , further comprising:
using a separate dense neural network for each finger.
28. The method as in claim 19 , further comprising:
using a random pose generator to set a pose of a 3D model for each sample in a synthetic source domain dataset.
29. The method as in claim 27 , further comprising:
adding Gaussian noise to pose key-point inputs while training the second neural network.
30. The method as in claim 23 , wherein the domain discriminator neural network uses global average pooling so that only small-scale domain-invariant features are learned.