Extrapolating lighting conditions from a single digital image

1. In a digital medium environment for analyzing and editing digital images, a computer-implemented method of estimating illumination from digital images, the computer-implemented method comprising:

estimating a set of lighting parameters from a single low-dynamic range digital image by feeding the single low-dynamic range digital image into a neural network trained using a plurality of pairs of panoramic images and limited-field-of-view images to output lighting parameters from a single input image;

generating, based on the set of lighting parameters, high-dynamic range lighting conditions for the single low-dynamic range digital image; and

generating a modified digital image from the single low-dynamic range digital image using the generated high-dynamic range lighting conditions.

2. The computer-implemented method as recited in claim 1 , wherein estimating the set of lighting parameters from the single low-dynamic range digital image comprises estimating a sun position for the single low-dynamic range digital image.

3. The computer-implemented method as recited in claim 2 , wherein estimating the sun position for the single low-dynamic range digital image comprises generating a probability distribution over likely sun positions.

4. The computer-implemented method as recited in claim 2 , wherein the single low-dynamic range digital image is devoid of a captured sun.

5. The computer-implemented method as recited in claim 2 , wherein estimating the set of lighting parameters from the single low-dynamic range digital image comprises estimating an atmospheric turbidity and an exposure.

6. The computer-implemented method as recited in claim 1 , further comprising:

estimating, using the neural network and from the single low-dynamic range digital image, one or more camera parameters associated with an image capturing device used to capture the single low-dynamic range digital image;

calibrating a virtual camera with the estimated one or more camera parameters; and

inserting an object into the modified digital image using the calibrated virtual camera.

7. The computer-implemented method as recited in claim 1 , further comprising training the neural network by:

analyzing a set of low-dynamic range panoramic images to determine a plurality of ground truth lighting parameters;

feeding into the neural network a limited-field-of-view image from each low-dynamic range panoramic image to estimate lighting parameters for the limited-field-of-view image;

comparing, for each low-dynamic range panoramic image, the ground truth lighting parameters with the estimated lighting parameters of the corresponding limited-field-of-view image; and

minimizing a loss between the estimated lighting parameters and the ground truth lighting parameters.

8. The computer-implemented method as recited in claim 1 , wherein generating, based on the estimated set of lighting parameters, the high-dynamic range lighting conditions for the single low-dynamic range digital image comprises generating a high-dynamic range outdoor environment map.

9. The computer-implemented method as recited in claim 1 , wherein generating the modified digital image from the single low-dynamic range digital image using the generated high dynamic lighting conditions comprises inserting an object into the single low-dynamic range digital image, the object being illuminated according to the generated high-dynamic range lighting conditions.

10. A system for estimating illumination from digital images, comprising:

a memory comprising a set of panoramic images each depicting a view of a sun; and

a computing device comprising instructions thereon that, when executed by at least one processor, cause the computing device to:

train a neural network to output lighting parameters from a single input image by:

analyzing the set of panoramic images to determine a plurality of ground truth image lighting parameters; and

iteratively:

feeding, into the neural network, a limited-field-of-view image from a panoramic image to determine a plurality of estimated lighting parameters, each the limited-field-of-view image comprising a sub-portion of one of the panoramic images;

comparing, for each panoramic image, the ground truth lighting parameters with the estimated lighting parameters of the corresponding the limited-field-of-view image; and

minimizing a loss between the estimated lighting parameters and the ground truth lighting parameters.

11. The system as recited in claim 10 , wherein the panoramic images comprise low-dynamic range spherical 360 degree images.

12. The system as recited in claim 10 , wherein feeding, into the neural network, the limited-field-of-view image from the panoramic image to determine a plurality of estimated lighting parameters comprises feeding multiple limited view portions of each panoramic image into the neural network.

13. The system as recited in claim 10 , wherein analyzing the set of panoramic images to determine the plurality of ground truth image lighting parameters and feeding, into the neural network, the limited-field-of-view image from the panoramic image to determine the plurality of estimated lighting parameters each comprise determining a sun position and sky parameters, the sky parameters comprising an exposure and a ground truth turbidity.

14. The system as recited in claim 13 , wherein the neural network comprises a first head for estimating the sun position and a second head for estimating the sky parameters.

15. The system as recited in claim 14 , wherein minimizing the loss between the estimated lighting parameters and the ground truth lighting parameters comprises:

determining a summed loss by adding a sun position loss and a sky parameters loss; and

minimizing the summed loss via stochastic gradient descent.

16. The system as recited in claim 10 , wherein analyzing the set of panoramic images to determine the plurality of ground truth lighting parameters comprises applying a sky model algorithm to determine the ground truth lighting parameters.

17. The system as recited in claim 10 , further comprising instructions that, when executed by the at least one processor, cause the computing device to:

extract a feature representation that captures illumination information of the single input image from a layer of the neural network; and

use the feature representation to analyze, categorize, or search additional digital images by comparing the feature representation to feature representations of the additional digital images.

18. In a digital medium environment for analyzing and editing digital images, a computer-implemented method of estimating illumination from digital images, the computer-implemented method comprising:

receiving a single low-dynamic range digital image;

receiving a request to modify the single low-dynamic range digital image;

a step for generating high-dynamic range lighting conditions from the single low-dynamic range digital image; and

generating a modified digital image from the single low-dynamic range digital image using the high-dynamic range lighting conditions.

19. The computer-implemented method as recited in claim 18 , wherein the step for determining high-dynamic range lighting conditions from the single low-dynamic range digital image comprises estimating a set of camera parameters associated with an image capturing device used to capture the single low-dynamic range digital image.

20. The computer-implemented method as recited in claim 18 , wherein the step for generating determining high-dynamic range lighting conditions from the single low-dynamic range digital image comprises training a neutral network to estimate high-dynamic range lighting conditions from low-dynamic range digital images.