Machine learning models for image injection that compensate for artifacts
A method including an image is received, a modified image is generated based on the image using a trained machine learning model, a rendered image is generated based on the modified image using an image injection pipeline, and the rendered image is displayed on a display of a wearable device.
1 . A method comprising:
receiving a code representation of an image from a companion device communicatively coupled to a wearable device, the code representation being generated using a first trained model;
generating a modified image based on the code representation using a second trained model;
generating a rendered image based on the modified image using an image injection pipeline; and
displaying the rendered image on a display of the wearable device.
2 . The method of claim 1 , wherein,
the image injection pipeline introduces artifacts to the rendered image, and
the second trained model is trained to compensate for the artifacts introduced by the image injection pipeline.
3 . The method of claim 1 , further comprising:
before generating the rendered image, communicating the modified image from the companion device to the wearable device.
4 . The method of claim 1 , wherein the image is received from the companion device.
5 . The method of claim 1 , wherein the image is generated by the wearable device.
6 . The method of claim 1 , wherein the second trained model is an autoencoder.
7 . The method of claim 1 , wherein
the image injection pipeline includes a light engine assembly and a see-through stack assembly, and
the second trained model is trained to compensate for artifacts introduced by at least one of the light engine assembly or the see-through stack assembly.
8 . The method of claim 7 , wherein the artifacts are based on at least one of luminance, luminance uniformity, color, color uniformity, contrast, light direction, or a pixel pattern of the image.
9 . The method of claim 1 , wherein
the second trained model is trained to generate the modified image by adding inversion artifacts to an image decoded from the code representation, and
the inversion artifacts are configured to cancel artifacts introduced by the image injection pipeline.
10 . The method of claim 9 , wherein the artifacts are based on at least one of luminance, luminance uniformity, color, color uniformity, contrast, light direction, or a pixel pattern of the image.
11 . A method comprising:
receiving a ground-truth image;
generating a modified image based on the ground-truth image using a model;
generating a rendered image based on the modified image using an image injection pipeline; and
training the model based on a comparison of the rendered image to the ground-truth image, wherein the training of the model compensates for artifacts introduced by the image injection pipeline by causing inversion effects that are opposite of the artifacts introduced by the image injection pipeline.
12 . The method of claim 11 , further comprising:
generating a loss based on the comparison of the rendered image to the ground-truth image; and
modifying a weight associated with the model based on the loss, wherein,
the artifacts introduced by the image injection pipeline vary portions of the ground-truth image, and
at least one of the loss and the modified weight is varied based on the artifacts introduced by the image injection pipeline varying portions of the ground-truth image.
13 . The method of claim 12 , wherein,
the loss is a mean squared loss, and
calculating the mean squared loss varies based on the artifacts introduced by a rendering pipeline varying portions of the ground-truth image.
14 . The method of claim 12 , wherein,
the loss is a cost function that uses a first cost and a second cost greater than the first cost,
a first type of artifact has the first cost, and
a second type of artifact has the second cost.
15 . The method of claim 11 , wherein the model is an autoencoder.
16 . A non-transitory computer-readable storage medium comprising instructions stored thereon that, when executed by at least one processor, are configured to cause a computing system to:
receive a code representation of an image from a companion device communicatively coupled to a wearable device, the code representation being generated using a first trained model;
generate a modified image based on the code representation using a second trained model;
generate a rendered image based on the modified image using an image injection pipeline; and
display the rendered image on a display of the wearable device.
17 . The non-transitory computer-readable storage medium of claim 16 , wherein the instructions further cause the computing system to:
before generating the rendered image, communicate the modified image from the companion device to the wearable device.
18 . The non-transitory computer-readable storage medium of claim 16 , wherein the second trained model is an autoencoder.
19 . The non-transitory computer-readable storage medium of claim 16 , wherein
the image injection pipeline introduces artifacts to the rendered image, and
the second trained model is trained to compensate for the artifacts introduced by the image injection pipeline.
20 . The non-transitory computer-readable storage medium of claim 16 , wherein the image is generated by the wearable device.