Classification for image creation
Implementations disclosed herein provide systems and methods that use classification-based machine learning to generate perceptually-plausible content for a missing part (e.g., some or all) of an image. The machine learning model may be trained to generate content for the missing part that appears plausible by learning to generate content that cannot be distinguished from real image content, for example, using adversarial loss-based training. To generate the content, a probabilistic classifier may be used to select color attribute values (e.g., RGB values) for each pixel of the missing part of the image. To do so, a pixel color attribute is segmented into a number of bins (e.g., value ranges) that are used as classes. The classifier determines probabilities for each of the bins of a color attribute for each pixel and generates the content by selecting the bin having the highest probability for each color attribute for each pixel.
1. A method comprising:
at a mobile electronic device having a processor:
receiving input data corresponding to an image of an environment, wherein the image has a missing part comprising pixels for content that is undefined by the input data; and
generating content for the missing part of the image by inputting the input data to a machine-learning model, wherein a color attribute is quantized into bins and the machine-learning model is configured to generate the content by performing a classification for each pixel of the missing part for the bins and selecting a color attribute for each pixel by selecting one of the bins based on the classification.
2. The method of claim 1 , wherein the missing part comprises pixels for which color values are unknown prior to generating the content.
3. The method of claim 1 , wherein the input data comprises an image comprising pixels for which color values are unknown prior to generating the content.
4. The method of claim 1 , wherein the input data comprises an equirectangular representation or a cube map representing a 360 degree view of a three dimensional (3D) environment.
5. The method of claim 1 , wherein the input data comprises a vector or statement identifying attributes of the image or portion of the image to be generated.
6. The method of claim 1 , wherein the input data comprises a mask identifying which pixels of the image are to be generated.
7. The method of claim 1 , wherein the missing part is less than all of the image, wherein the image has a defined part comprising pixels for content that are defined by the input data.
8. The method of claim 1 , wherein the missing part is all of the image and all of the content of the image is synthesized by the machine-learning model.
9. The method of claim 1 , wherein the machine-learning model is a classification-based neural network.
10. The method of claim 1 , wherein the classification provides a probability for each of the bins and a bin having the highest probability value for each pixel is selected.
11. The method of claim 1 , wherein
each of multiple color channels is quantized into bins; and
the machine-learning model is configured to generate the content by performing a classification for each pixel of the missing part for the bins of each color channel and selecting the color attribute for each pixel for each channel by selecting one of the bins based on the classification.
12. The method of claim 1 , wherein the machine-learning model is trained using a cross entropy-based classification loss.
13. The method of claim 1 , wherein machine-learning model is trained using an adversarial loss that penalizes generation of content that an adversarial network can distinguish from real content.
14. The method of claim 1 , wherein machine-learning model is trained using a perceptual loss.
15. The method of claim 1 , wherein the content is generated to fill a hole in the image.
16. The method of claim 1 , wherein the content is generated to synthesize an object based on input data that identifies an object appearance attribute, an object class, or incident light.
17. The method of claim 1 , wherein the content is generated to synthesize a translucent object.
18. The method of claim 1 , wherein the content is generated to convert a sketch to an image.
19. The method of claim 1 , wherein the content is generated to fill in missing parts of a computer-generated reality (CGR) environment.
20. The method of claim 1 , wherein selecting the color attribute for each pixel comprises selecting a color value for a pixel that had an unknown color value prior to generating the content.
21. A system comprising:
a non-transitory computer-readable storage medium; and
one or more processors coupled to the non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium comprises program instructions that, when executed on the one or more processors, cause the system to perform operations comprising:
receiving input data corresponding to an image of an environment, wherein the image has a missing part comprising pixels for content that is undefined; and
generating content for the missing part of the image by inputting the input data to a machine-learning model, wherein a color attribute is quantized into bins and the machine-learning model is configured to generate the content by performing a classification for each pixel of the missing part for the bins and selecting a color attribute for each pixel by selecting one of the bins based on the classification.
22. The method of claim 20 , wherein the color value corresponds to a value identifying a color or a value for a color channel.