Unsupervised real-to-virtual domain unification for end-to-end highway driving
An unsupervised real to virtual domain unification model for highway driving, or DU-drive, employs a conditional generative adversarial network to transform driving images in a real domain to their canonical representations in the virtual domain, from which vehicle control commands are predicted. In the case where there are multiple real datasets, a real-to-virtual generator may be independently trained for each real domain and a global predictor could be trained with data from multiple real domains. Qualitative experiment results show this model can effectively transform real images to the virtual domain while only keeping the minimal sufficient information, and quantitative results verify that such canonical representation can eliminate domain shift and boost the performance of control command prediction task.
1. A method of commanding an operation of an autonomous device, the method comprising:
obtaining, by a domain unification (DU)-drive model, one or more real operating images associated with operation of the autonomous device;
transforming, by the DU-drive model, each of the one or more real operating images to its corresponding canonical representation, wherein the canonical representation of a real operating image from the one or more real operating images, is a pixel-level fake virtual image of the real operating image, and wherein the transforming each of the one or more real operating images comprises:
mapping each of the one or more real operating images to a corresponding true virtual image sampled from a database of true virtual images having true virtual images that are annotated with corresponding ground-truth operating parameters of the autonomous device; and
predicting, by the DU-drive model, an operating parameter for the autonomous device based on the one or more canonical representations by training the DU-drive model to minimize a mean square loss between the predicted operating parameter and the ground truth operating parameter of the autonomous device, wherein the one or more real operating images comprise predictive information and background information and its corresponding canonical representation contains at least some of the predictive information from which to predict the operating parameter, wherein the predictive information corresponds to lane markings on a road, wherein the DU-drive model predicts the operating parameter to be a steering angle based on the lane markings.
2. The method of claim 1 , wherein the one or more real operating images are obtained from a camera associated with the autonomous device.
3. The method of claim 1 , wherein the background information corresponds to one or more of objects in a horizon or off to either side of the road.
4. The method of claim 1 , wherein the true virtual images in the database are related to an operation of a computer simulated device and are generated using computer graphics technology.
5. The method of claim 1 , wherein predicting the operating parameter comprises predicting a moving angle of the autonomous device.
6. The method of claim 1 , further comprising:
controlling one or more control systems of the autonomous device based on the predicted operating parameter.
7. The method of claim 1 , further comprising:
generating the canonical representation of a real operating image of the one or more real operating images; and
predicting the operating parameter based on the generated canonical representation.
8. The method of claim 7 , wherein the generation of the canonical representation and the prediction of the operating parameter comprises using a generative adversarial neural network (GAN) using a first database of real source images annotated with corresponding ground-truth operating parameters obtained from a real operating experience, wherein the real source images are images of real-world scenes related to the operation of other autonomous devices, and a second database of virtual source images annotated with corresponding ground-truth operating parameters obtained from the same operating experience, wherein the virtual source images are generated using computer graphics technology.
9. The method of claim 8 , wherein generating the canonical representation of the real operating image by using the GAN comprises:
generating fake virtual images for fooling a process of discrimination performed to discriminate fake virtual images from true virtual images, and
correctly discriminating the generated fake virtual images from true virtual images.
10. An apparatus for commanding an operation of an autonomous device, the apparatus comprising:
a memory; and
a processor coupled to the memory and configured to:
obtain one or more real operating images associated with operation of the autonomous device, and
transform each of the one or more real operating images to its corresponding canonical representation, wherein the canonical representation of a real operating image from the one or more real operating images, is a pixel-level fake virtual image of the real operating image, and wherein the transforming each of the one or more real operating images comprises:
mapping each of the one or more real operating images to a corresponding true virtual image sampled from a database of true virtual images having true virtual images that are annotated with corresponding ground-truth operating parameters of the autonomous device; and
predict an operating parameter for the autonomous device based on the one or more canonical representations by training the processor to minimize a mean square loss between the predicted operating parameter and the ground-truth operating parameter of the autonomous device, wherein the one or more real operating images comprise predictive information and background information and its corresponding canonical representation contains at least some of the predictive information from which to predict the operating parameter, wherein the predictive information corresponds to lane markings on a road, wherein the processor predicts the operating parameter to be a steering angle based on the lane markings.
11. The apparatus of claim 10 , wherein the one or more real operating images are obtained from a camera associated with the autonomous device.
12. The apparatus of claim 10 , wherein the background information comprises one or more of objects in a horizon or off to either side of the road.
13. The apparatus of claim 10 , wherein the true virtual images in the database are related to an operation of a computer simulated device and are generated using computer graphics technology.
14. The apparatus of claim 10 , wherein the processor is further configured to output a signal that controls one or more control systems of the autonomous device based on the predicted operating parameter.
15. The apparatus of claim 10 , wherein the autonomous device is one of an autonomous or semi-autonomous car, drone, and robot.
16. The apparatus of claim 10 , wherein the processor is configured to:
generate the canonical representation of a real operating image of the one or more real operating images; and
predict the operating parameter based on the generated canonical representation.
17. The apparatus of claim 16 , wherein for generating the canonical representation, the processor is configured to map the real operating image to the canonical representation of the real operating image that contains the at least some of the predictive information from which to predict the operating parameter.
18. The apparatus of claim 16 , wherein for generating the canonical representation and predicting the operating parameter, the processor is configured to use a generative adversarial neural network (GAN) using a first database of real source images annotated with corresponding ground-truth operating parameters obtained from a real operating experience, wherein the real source images are images of real-world scenes related to the operation of other autonomous devices, and a second database of virtual source images annotated with corresponding ground-truth operating parameters obtained from the same operating experience, wherein the virtual source images are generated using computer graphics technology.
19. The apparatus of claim 18 , wherein generating the canonical representation of the real operating image by using the GAN comprises:
generating fake virtual images for fooling a process of discrimination performed to discriminate fake virtual images from true virtual images, and
correctly discriminating the generated fake virtual images from the true virtual images.