Model-aware synthetic image generation

1. A computer implemented method comprising:

obtaining a first deep neural network (DNN) model trained on labeled real image data for a downstream vision task;

obtaining a second DNN model trained on synthetic images created with random image parameter values for the downstream vision task;

obtaining a third DNN model trained on the labeled real image data and the synthetic images for the downstream vision task;

performing a forward pass execution of each model to generate a loss;

backpropagating the loss to modify parameter values; and

iterating the forward pass execution and backpropagating with images generated by the modified parameter values to jointly train the models and optimize parameters.

2. The method of claim 1 wherein the parameters comprise content generation parameters and rendering parameters.

3. The method of claim 2 wherein the parameters are modified by knobs that are mapped into a differentiable continuous space.

4. The method of claim 3 and further comprising mapping knobs corresponding to the optimized parameters back to discrete space.

5. The method of claim 4 and further comprising using the discrete space mapped knobs to generate additional synthetic images.

6. The method of claim 4 and further comprising training the DNN models using the generated additional synthetic images.

7. The method of claim 6 and further comprising continuing to train the DNN models with the generated additional synthetic images until a pre-selected model performance threshold is reached.

8. The method of claim 4 wherein the synthetic images are generated by transforming the real image data.

9. The method of claim 1 wherein the loss is calculated by a loss function based on a combination of downstream vision task metrics and pixel similarity between real and synthetic images.

10. The method of claim 9 wherein the loss function is further based on one or more of a model execution path, a neuron firing pattern, or a neuron coverage, one or more of a content or mask constraint, and a model feature similarity across the models.

11. The method of claim 1 wherein model parameters are fixed while backpropagating the loss.

12. A machine-readable storage device having instructions for execution by a processor of a machine to cause the processor to perform operations to perform a method, the operations comprising:

obtaining first deep neural network (DNN) model trained on labeled real image data for a downstream vision task;

obtaining a second DNN model trained on synthetic images created with random image parameter values for the downstream vision task: obtaining a third DNN model trained on the labeled real image data and the synthetic images for the downstream vision task;

performing a forward pass execution of each model to generate a loss;

backpropagating the loss to modify parameter values; and

iterating the forward pass execution and backpropagating with images generated by the modified parameter values to jointly train the models and optimize parameters.

13. The device of claim 12 wherein the parameters comprise content generation parameters and rendering parameters that are modified by knobs that are mapped into a differentiable continuous space.

14. The device of claim 13 wherein the operations further comprise mapping knobs corresponding to the optimized parameters back to discrete space.

15. The device of claim 14 wherein the operations further comprise using the discrete space mapped knobs to generate additional synthetic images.

16. The device of claim 14 wherein the operations further comprise training the DNN models using the generated additional synthetic images and continuing to train the DNN models with the generated additional synthetic images until a pre-selected model performance threshold is reached.

17. The device of claim 12 wherein the loss is calculated by a loss function based on a combination of downstream vision task metrics and pixel similarity between real and synthetic images.

18. The device of claim 17 wherein the loss function is further based on one or more of a model execution path, a neuron firing pattern, or a neuron coverage, one or more of a content or mask constraint, and a model feature similarity across the models and wherein model parameters are fixed while backpropagating the loss.

19. A device comprising:

a processor; and

a memory device coupled to the processor and having a program stored thereon for execution by the processor to perform operations comprising:

obtaining first deep neural network (DNN) model trained on labeled real image data for a downstream vision task;

obtaining a second DNN model trained on synthetic images created with random image parameter values for the downstream vision task;

obtaining a third DNN model trained on the labeled real image data and the synthetic images for the downstream vision task;

performing a forward pass execution of each model to generate a loss; backpropagating the loss to modify parameter values; and

iterating the forward pass execution and backpropagating with images generated by the modified parameter values to jointly train the models and optimize parameters.

20. The device of claim 19 wherein the parameters comprise content generation parameters and rendering parameters that are modified by knobs that are mapped into a differentiable continuous space and wherein the operations further comprise training the DNN models using generated additional synthetic images and continuing to train the DNN models with the generated additional synthetic images until a pre-selected model performance threshold is reached, wherein the loss is calculated by a loss function based on a combination of downstream vision task metrics and pixel similarity between real and synthetic images.