Generating quasi-realistic synthetic training data for use with machine learning models
Implementations are described herein for automatically generating quasi-realistic synthetic training images that are usable as training data for training machine learning models to perceive various types of plant traits in digital images. In various implementations, multiple labeled simulated images may be generated, each depicting simulated and labeled instance(s) of a plant having a targeted plant trait. In some implementations, the generating may include stochastically selecting features of the simulated instances of plants from a collection of plant assets associated with the targeted plant trait. The collection of plant assets may be obtained from ground truth digital image(s). In some implementations, the ground truth digital image(s) may depict real-life instances of plants having the target plant trait. The plurality of labeled simulated images may be processed using a trained generator model to generate a plurality of quasi-realistic synthetic training images, each depicting quasi-realistic and labeled instance(s) of the targeted plant trait.
1. A method implemented using one or more processors, the method comprising:
generating a plurality of labeled simulated images, each labeled simulated image depicting one or more simulated and labeled instances of a plant having a targeted plant trait,
wherein the generating includes stochastically selecting features of the simulated instances of plants from a collection of plant assets associated with the targeted plant trait, wherein the collection of plant assets are obtained from one or more ground truth digital images captured previously by one or more vision sensors, and
wherein the one or more ground truth digital images depict real-life instances of plants having the target plant trait; and
subsequent to generating the plurality of labeled simulated images, translating the plurality of labeled simulated images from a simulated domain to a quasi-realistic domain using a trained conditional generator model of a generative adversarial network, wherein the translating includes preserving the labels from the plurality of labeled simulated images and generating a plurality of quasi-realistic synthetic training images based at least in part on random noise and subject to one or more user-provided constraints, including a shape defined by user input, wherein each quasi-realistic synthetic training image depicts one or more quasi-realistic and labeled instances of the targeted plant trait.
2. The method of claim 1 , wherein the generator model is trained at least in part based on one or more of the ground truth digital images.
3. The method of claim 1 , further comprising analyzing the one or more ground truth digital images to segment depicted plant assets for inclusion in the collection of plant assets.
4. The method of claim 3 , wherein the depicted plant assets include one or more features of leaves of plant(s) having the targeted plant trait.
5. The method of claim 1 , wherein the plurality of labeled simulated images are labeled with pixel-wise annotations or bounding shapes.
6. The method of claim 1 , wherein the stochastically selecting is based on different weights assigned to different plant assets of the collection.
7. The method of claim 6 , wherein the different weights are selected based on one or more environmental conditions observed or predicted in a selected geographic area.
8. The method of claim 1 , wherein the generative adversarial network comprises a cycle generative adversarial network (cycleGAN).
9. A system comprising one or more processors and memory storing instructions that, in response to execution of the instructions by the one or more processors, cause the one or more processors to:
generate a plurality of labeled simulated images, each labeled simulated image depicting one or more simulated and labeled instances of a plant having a targeted plant trait,
wherein the generation includes stochastically selection of features of the simulated instances of plants from a collection of plant assets associated with the targeted plant trait, wherein the collection of plant assets are obtained from one or more ground truth digital images captured previously by one or more vision sensors, and
wherein the one or more ground truth digital images depict real-life instances of plants having the target plant trait; and
once the plurality of labeled simulated images are generated, translate the plurality of labeled simulated images from a simulated domain to a quasi-realistic domain using a trained conditional generator model of a generative adversarial network, wherein the translation includes preservation of the labels from the plurality of labeled simulated images and generation of a plurality of quasi-realistic synthetic training images based at least in part on random noise and subject to one or more user-provided constraints, including a shape defined by user input, wherein each quasi-realistic synthetic training image depicts one or more quasi-realistic and labeled instances of the targeted plant trait.
10. The system of claim 9 , wherein the generator model is trained at least in part based on one or more of the ground truth digital images.
11. The system of claim 9 , further comprising instructions to analyze the one or more ground truth digital images to segment depicted plant assets for inclusion in the collection of plant assets.
12. The system of claim 11 , wherein the depicted plant assets include one or more features of leaves of plant(s) having the targeted plant trait.
13. The system of claim 9 , wherein the plurality of labeled simulated images are labeled with pixel-wise annotations or bounding shapes.
14. The system of claim 9 , wherein the stochastically selecting is based on different weights assigned to different plant assets of the collection.
15. The system of claim 14 , wherein the different weights are selected based on one or more environmental conditions observed or predicted in a selected geographic area.
16. The system of claim 9 , wherein the generative adversarial network comprises a cycle generative adversarial network (cycleGAN).
17. At least one non-transitory computer-readable medium comprising instructions that, in response to execution of the instructions by one or more processors, cause the one or more processors to perform the following operations:
generating a plurality of labeled simulated images, each labeled simulated image depicting one or more simulated and labeled instances of a plant having a targeted plant trait,
wherein the generating includes stochastically selecting features of the simulated instances of plants from a collection of plant assets associated with the targeted plant trait, wherein the collection of plant assets are obtained from one or more ground truth digital images captured previously by one or more vision sensors, and
wherein the one or more ground truth digital images depict real-life instances of plants having the target plant trait; and
subsequent to the generating, translating the plurality of labeled simulated images from a simulated domain to a quasi-realistic domain using a trained conditional generator model of a generative adversarial network, wherein the translating includes preserving the labels from the plurality of labeled simulated images and generating a plurality of quasi-realistic synthetic training images based at least in part on random noise and subject to one or more user-provided constraints, including a shape defined by user input, wherein each quasi-realistic synthetic training image depicts one or more quasi-realistic and labeled instances of the targeted plant trait.
18. The at least one non-transitory computer-readable medium of claim 17 , wherein the generator model is trained at least in part based on one or more of the ground truth digital images.