Data augmentation for local feature detector and descriptor learning using appearance transform
Providing localization data includes obtaining a first image of a scene associated with a first condition, determining one or more target conditions, and applying an appearance transfer network to the first image to obtain one or more synthesized images comprising the scene, wherein the scene is associated with the one or more target conditions in the synthesized image. A first patch is selected from the first image, wherein the first patch comprises a keypoint, and an image location is determined for the first patch. Then one or more additional patches can be obtained using the synthesized images and the image location. A descriptor network may be trained to provide localization data based on the first patch and the one or more additional patches.
1. A non-transitory computer readable medium comprising computer readable instructions executable by one or more processors to:
obtain a first image of a scene associated with a first condition;
determine one or more target conditions;
apply an appearance transfer network to the first image to obtain one or more synthesized images comprising the scene, wherein the scene is associated with the one or more target conditions in the synthesized image;
select a first patch from the first image, wherein the first patch comprises a keypoint;
determine an image location of the first patch;
obtain a one or more additional patches from the one or more synthesized images based on the determined image location; and
train a descriptor network to provide localization data based on the first patch and the one or more additional patches.
2. The non-transitory computer readable medium of claim 1 , wherein the appearance transfer network is trained to generate an output image according to an input image and a target condition,
wherein the appearance transfer network is trained based on images of a scene associated with a plurality of conditions.
3. The non-transitory computer readable medium of claim 2 , wherein the images of the scene associated with a plurality of conditions comprises:
determining a plurality of sequences of images, wherein each sequence is associated with a unique condition of the plurality of conditions;
determining a trajectory associated with each of the plurality of sequences; and
align the images of the scene based on the determined trajectories.
4. The non-transitory computer readable medium of claim 1 , wherein the computer readable instructions to obtain a one or more additional patches further comprises computer readable instructions to:
determine a plurality of pixel locations comprising the image location of the first patch in the first image; and
for each synthesized image, obtain a set of pixels in the synthesized image at the plurality of pixel locations.
5. The non-transitory computer readable medium of claim 1 , wherein the localization data comprises condition-invariant descriptors.
6. The non-transitory computer readable medium of claim 1 , wherein the target conditions are associated with variations of illumination of the scene.
7. The non-transitory computer readable medium of claim 1 , wherein the target conditions are associated with variations of environmental conditions of the scene.
8. A system for providing localization data, comprising:
one or more processors; and
one or more computer readable medium comprising computer readable instructions executable by one or more processors to:
obtain a first image of a scene associated with a first condition;
determine one or more target conditions;
apply an appearance transfer network to the first image to obtain one or more synthesized images comprising the scene, wherein the scene is associated with the one or more target conditions in the synthesized image;
select a first patch from the first image, wherein the first patch comprises a keypoint;
determine an image location of the first patch;
obtain a one or more additional patches from the one or more synthesized images based on the determined image location; and
train a descriptor network to provide localization data based on the first patch and the one or more additional patches.
9. The system of claim 8 , wherein the appearance transfer network is trained to generate an output image according to an input image and a target condition,
wherein the appearance transfer network is trained based on images of a scene associated with a plurality of conditions.
10. The system of claim 9 , wherein the images of the scene associated with a plurality of conditions comprises:
determining a plurality of sequences of images, wherein each sequence is associated with a unique condition of the plurality of conditions;
determining a trajectory associated with each of the plurality of sequences; and
align the images of the scene based on the determined trajectories.
11. The system of claim 8 , wherein the computer readable instructions to obtain a one or more additional patches further comprises computer readable instructions to:
determine a plurality of pixel locations comprising the image location of the first patch in the first image; and
for each synthesized image, obtain a set of pixels in the synthesized image at the plurality of pixel locations.
12. The system of claim 8 , wherein the localization data comprises condition-invariant descriptors.
13. The system of claim 8 , wherein the target conditions are associated with variations of illumination of the scene.
14. The system of claim 8 , wherein the target conditions are associated with variations of environmental conditions of the scene.
15. A method for providing localization data, comprising:
obtaining a first image of a scene associated with a first condition;
determining one or more target conditions;
applying an appearance transfer network to the first image to obtain one or more synthesized images comprising the scene, wherein the scene is associated with the one or more target conditions in the synthesized image;
selecting a first patch from the first image, wherein the first patch comprises a keypoint;
determining an image location of the first patch;
obtaining a one or more additional patches from the one or more synthesized images based on the determined image location; and
training a descriptor network to provide localization data based on the first patch and the one or more additional patches.
16. The method of claim 15 , wherein the appearance transfer network is trained to generate an output image according to an input image and a target condition,
wherein the appearance transfer network is trained based on images of a scene associated with a plurality of conditions.
17. The method of claim 15 , wherein obtaining one or more additional patches further comprises:
determining a plurality of pixel locations comprising the image location of the first patch in the first image; and
for each synthesized image, obtaining a set of pixels in the synthesized image at the plurality of pixel locations.
18. The method of claim 15 , wherein the localization data comprises condition-invariant descriptors.
19. The method of claim 15 , wherein the target conditions are associated with variations of illumination of the scene.
20. The method of claim 15 , wherein the target conditions are associated with variations of environmental conditions of the scene.