Synthetic-to-realistic image conversion using generative adversarial network (GAN) or other machine learning model

An apparatus comprising:

• at least one memory configured to store:
• training data comprising first image pairs, each of the first image pairs comprising (i) a first training image and (ii) a first ground truth image;
• and
• additional training data comprising second image pairs, each of the second image pairs comprising (i) a second training image and (ii) a second ground truth image, wherein at least some of the images in the second image pairs are less aligned or of lower quality than at least some of the images in the first image pairs;
• and
• at least one processing device configured to:
• train a machine learning model to generate realistic images using the first image pairs;
• and
• continue to train the machine learning model to generate the realistic images using the second image pairs;
• wherein, to train the machine learning model and to continue to train the machine learning model, the at least one processing device is configured to use a loss function;
• and
• wherein the loss function is based on (i) a generative adversarial network loss and (ii) a feature matching loss.

A method comprising:

• obtaining one or more synthetic images of an environment;
• generating one or more realistic images of the environment based on the one or more synthetic images using a trained machine learning model;
• obtaining one or more actual images of the environment;
• and
• determining at least one characteristic of a flight vehicle based on the one or more realistic images of the environment and the one or more actual images of the environment;
• wherein the trained machine learning model is trained by:
• training a machine learning model to generate realistic images using first image pairs, each of the first image pairs comprising (i) a first training image and (ii) a first ground truth image;
• and
• continuing to train the machine learning model to generate the realistic images using second image pairs, each of the second image pairs comprising (i) a second training image and (ii) a second ground truth image, wherein at least some of the images in the second image pairs are less aligned or of lower quality than at least some of the images in the first image pairs;
• wherein training the machine learning model and continuing to train the machine learning model comprise using a loss function;
• and
• wherein the loss function is based on (i) a generative adversarial network loss and (ii) a feature matching loss.