Synthesizing training data for training a change detection model
Synthesizing training data for training a change detection model includes receiving a patched image comprising a background image with a patch pasted into the background image. It further includes synthesizing a harmonized patched image at least in part by harmonizing the patched image using a machine learning model trained on the background image. It further includes providing as output a synthetic training sample usable to train a change detection model. The synthetic training sample includes a reference image, at least a portion of the harmonized patched image, and a corresponding mask.
1. A system, comprising:
one or more processors configured to:
receive a patched image comprising a background image with a patch pasted into the background image, wherein the patch is sampled from an area surrounding the background image;
synthesize a harmonized patched image at least in part by harmonizing the patched image comprising the patch that is sampled from the area surrounding the background image and pasted into the background image, wherein the harmonizing comprises using a harmonization model trained on the background image; and
generate, based at least in part on the synthesizing using the harmonization model, a synthetic training sample comprising:
a reference image;
at least a portion of the harmonized patched image synthesized using the harmonization model; and
a corresponding target mask, wherein the target mask indicates a location in the harmonized patched image where the patch was inserted;
wherein, using the synthetic training sample that is generated based at least in part on the synthesizing using the harmonization model, a change detection model is trained at least in part by:
providing the reference image and the at least portion of the harmonized patched image as input to the change detection model;
comparing a change map outputted by the change detection model to the target mask included in the synthetic training sample; and
updating weights of the change detection model based on a difference determined between the change map outputted by the change detection model and the target mask included in the synthetic training sample that indicates the location where the patch was inserted in the harmonized patched image synthesized using the harmonization model; and
a memory coupled to the one or more processors and configured to provide the one or more processors with instructions.
2. The system recited in claim 1 , wherein the patch is randomly sampled.
3. The system recited in claim 1 , wherein the patch is sampled with a random size.
4. The system recited in claim 1 , wherein the patch is pasted at a random location within the background image.
5. The system recited in claim 1 , wherein training the harmonization model comprises training the harmonization model to reconstruct the background image.
6. The system recited in claim 1 , wherein the reference image comprises a portion of the background image corresponding to the at least portion of the harmonized patched image.
7. The system recited in claim 1 , wherein the background image comprises at least a portion of an image of a scene at a first time, and wherein the reference image comprises at least a portion of an image of the scene at a second time.
8. The system recited in claim 1 , wherein the one or more processors are configured to receive the background image, and wherein receiving the background image comprises performing cloud detection.
9. The system recited in claim 1 , wherein the harmonization model used to synthesize the harmonized patched image comprises a pyramidal generative adversarial network (GAN).
10. The system recited in claim 9 , wherein the patched image is provided as input to an intermediate stage of the pyramidal GAN.
11. The system recited in claim 1 , wherein the change detection model is configured to identify a change between a first image corresponding to a scene at a first time and a second image corresponding to the scene at a second time.
12. The system recited in claim 1 , wherein the change detection model takes as input a batch of images.
13. The system recited in claim 1 , wherein the change detection model takes as input multiband satellite imagery.
14. The system recited in claim 1 , wherein the change detection model comprises two encoder networks and a decoder network.
15. The system recited in claim 14 , wherein the two encoder networks are pre-trained, and wherein training the change detection model comprises:
fixing weights of the two encoder networks during training; and
training the decoder network.
16. A method, comprising:
receiving a patched image comprising a background image with a patch pasted into the background image, wherein the patch is sampled from an area surrounding the background image;
synthesizing a harmonized patched image at least in part by harmonizing the patched image comprising the patch that is sampled from the area surrounding the background image and pasted into the background image, wherein the harmonizing comprises using a harmonization model trained on the background image; and
generating, based at least in part on the synthesizing using the harmonization model, a synthetic training sample comprising:
a reference image;
at least a portion of the harmonized patched image synthesized using the harmonization model; and
a corresponding target mask, wherein the target mask indicates a location in the harmonized patched image where the patch was inserted;
wherein, using the synthetic training sample that is generated based at least in part on the synthesizing using the harmonization model, a change detection model is trained at least in part by:
providing the reference image and the at least portion of the harmonized patched image as input to the change detection model;
comparing a change map outputted by the change detection model to the target mask included in the synthetic training sample; and
updating weights of the change detection model based on a difference determined between the change map outputted by the change detection model and the target mask included in the synthetic training sample that indicates the location where the patch was inserted in the harmonized patched image synthesized using the harmonization model.
17. A computer program product embodied in a non-transitory computer readable storage medium and comprising computer instructions for:
receiving a patched image comprising a background image with a patch pasted into the background image, wherein the patch is sampled from an area surrounding the background image;
synthesizing a harmonized patched image at least in part by harmonizing the patched image comprising the patch that is sampled from the area surrounding the background image and pasted into the background image, wherein the harmonizing comprises using a harmonization model trained on the background image; and
generating, based at least in part on the synthesizing using the harmonization model, a synthetic training sample comprising:
a reference image;
at least a portion of the harmonized patched image synthesized using the harmonization model; and
a corresponding target mask, wherein the target mask indicates a location in the harmonized patched image where the patch was inserted;
wherein, using the synthetic training sample that is generated based at least in part on the synthesizing using the harmonization model, a change detection model is trained at least in part by:
providing the reference image and the at least portion of the harmonized patched image as input to the change detection model;
comparing a change map outputted by the change detection model to the target mask included in the synthetic training sample; and
updating weights of the change detection model based on a difference determined between the change map outputted by the change detection model and the target mask included in the synthetic training sample that indicates the location where the patch was inserted in the harmonized patched image synthesized using the harmonization model.