GAN-based data synthesis for semi-supervised learning of a radar sensor
Examples disclosed herein relate to a method for semi-supervised training of a radar system. The method includes training a first radar network of the radar system with a first set of radar object detection labels corresponding to a first set of radar data, training a generative adversarial network (GAN) with the trained first radar network, synthesizing a training data set for a second radar network of the radar system with the trained GAN, training a second radar network with the synthesized training data set, and generating a second set of radar object detection labels based on the training of the second radar network.
1. A method for semi-supervised training of a radar system, comprising:
training a first radar network of the radar system with a first set of radar object detection labels corresponding to a first set of radar data;
training a generative adversarial network (GAN) with the trained first radar network;
synthesizing a training data set for a second radar network of the radar system with the trained GAN;
training the second radar network with the synthesized training data set; and
generating a second set of radar object detection labels based on the training of the second radar network.
2. The method of claim 1 , wherein the first set of radar data is acquired via one or more of lidar or camera sensors of the radar system during an actual driving condition that occurs on a plurality of roads at various times with various weather patterns for one or more days.
3. The method of claim 1 , further comprising:
training one or more of lidar or camera sensors on actual driving conditions, wherein the training of the one or more of lidar or camera sensors occurs during the training of the first radar network of the radar system.
4. The method of claim 1 , wherein the synthesized training data set comprises a data set and corresponding labels that are generated during an inference from the first set of radar data and the first set of radar object detection labels.
5. The method of claim 1 , wherein the GAN is a neural network configured to distinguish real data from fake data, wherein the real data comprises the first set of radar data obtained during actual driving conditions acquired via one or more of lidar or camera sensors of the radar system.
6. The method of claim 5 , wherein the GAN comprises a combination of a generative network and a discriminative network, wherein the generative network is configured for generating new data instances based on the first set of radar data and the discriminative network is configured to distinguish the new data instances from the fake data.
7. The method of claim 1 , wherein the first radar network comprises a first beam steering radar and the second radar network comprises a second beam steering radar, wherein the first beam steering radar and the second beam steering radar have different characteristics in at least one of a configuration of parameters or a scan pattern.
8. The method of claim 7 , wherein the training of the second radar network by the GAN is performed prior to a deployment of the second beam steering radar.
9. A system for training a radar, comprising:
a first radar network that provides a first set of radar object detection labels corresponding to a first set of radar data;
a generative adversarial network (GAN) module trained with the first radar network;
a second radar network; and
the GAN module configured to train the second radar network using the first set of radar object detection labels and the first set of radar data.
10. The system of claim 9 , wherein the first set of radar data is acquired via one or more of lidar or camera sensors of the radar during an actual driving condition that occurs on a plurality of roads at various times with various weather patterns for one or more days.
11. The system of claim 9 , wherein the GAN module comprises a neural network configured to distinguish real data from fake data, wherein the real data comprises the first set of radar data obtained during actual driving conditions acquired via one or more of lidar or camera sensors of the radar.
12. The system of claim 11 , wherein the GAN module comprises a combination of a generative network and a discriminative network, wherein the generative network is configured for generating new data instances based on the first set of radar data and the discriminative network is configured to distinguish the new data instances from the fake data.
13. The system of claim 9 , wherein the GAN module is configured to synthesize a training set for the second radar network of the radar using the first set of radar object detection labels and the first set of radar data.
14. The system of claim 9 , wherein the first radar network comprises a first beam steering radar and the second radar network comprises a second beam steering radar, wherein the first beam steering radar and the second beam steering radar have different characteristics in at least one of a configuration of parameters or a scan pattern.
15. The system of claim 14 , wherein the second radar network is trained with the GAN module prior to a deployment of the second beam steering radar.
16. A non-transitory computer readable medium comprising computer executable instructions stored thereon to cause one or more processing units to:
train a first radar network of a radar system with a first set of radar object detection labels corresponding to a first set of radar data;
train a generative adversarial network (GAN) with the trained first radar network; synthesize a training data set for a second radar network of the radar system with the trained GAN;
train the second radar network with the synthesized training data set; and
generate a second set of radar object detection labels based on the training of the second radar network.
17. The non-transitory computer readable medium of claim 16 , wherein the first set of radar data is acquired via one or more of lidar or camera sensors of the radar system during an actual driving condition that occurs on a plurality of roads at various times with various weather patterns for one or more days.
18. The non-transitory computer readable medium of claim 16 , wherein the instructions further cause the one or more processing units to:
train one or more of lidar or camera sensors on actual driving conditions, wherein the training of the one or more of lidar or camera sensors occurs during the training of the first radar network of the radar system.
19. The non-transitory computer readable medium of claim 16 , wherein the GAN is a neural network configured to distinguish real data from fake data, wherein the GAN comprises a combination of a generative network and a discriminative network, wherein the generative network is configured for generating new data instances based on the first set of radar data and the discriminative network is configured to distinguish the new data instances from the fake data.
20. The non-transitory computer readable medium of claim 16 , wherein the first radar network comprises a first beam steering radar and the second radar network comprises a second beam steering radar, wherein the first beam steering radar and the second beam steering radar have different characteristics in at least one of a configuration of parameters or a scan pattern, and wherein the training of the second radar network by the GAN is performed prior to deployment of the second beam steering radar.