Generating refined segmentation masks based on uncertain pixels

1. A non-transitory computer-readable medium storing instructions thereon that, when executed by at least one processor, cause a computing device to:

identify a digital visual media item comprising a plurality of pixels, the digital visual media item depicting one or more objects; and

utilize a segmentation refinement neural network to:

generate an initial segmentation mask for the digital visual media item by determining whether the plurality of pixels correspond to the one or more objects by generating probabilities that the plurality of pixels correspond to the one or more objects;

determine, based on the initial segmentation mask, uncertain pixels, the uncertain pixels having an associated uncertainty that the uncertain pixels correspond to the one or more objects or do not correspond to the one or more objects; and

generate a refined segmentation mask for the digital visual media item by redetermining whether a set of uncertain pixels correspond to the one or more objects by generating updated probabilities that the plurality of pixels correspond to the one or more objects.

2. The non-transitory computer-readable medium of claim 1 , wherein the instructions, when executed by the at least one processor, cause the computing device to:

determine the uncertain pixels by generating an uncertainty map that identifies the uncertain pixels and further identifies certain pixels having an associated certainty that the certain pixels correspond to the one or more objects or do not correspond to the one or more objects; and

utilize the segmentation refinement neural network to generate the refined segmentation mask for the digital visual media item further based on feature values associated with the certain pixels identified by the uncertainty map.

3. The non-transitory computer-readable medium of claim 1 , further comprising instructions that, when executed by the at least one processor, cause the computing device to utilize the segmentation refinement neural network to:

generate one or more initial feature maps comprising low-level feature values corresponding to the digital visual media item; and

generate, based on the one or more initial feature maps, a final feature map comprising high-level feature values corresponding to the digital visual media item.

4. The non-transitory computer-readable medium of claim 3 , further comprising instructions that, when executed by the at least one processor, cause the computing device to utilize the segmentation refinement neural network to:

extract, from the low-level feature values, a subset of low-level feature values associated with the set of uncertain pixels; and

extract, from the high-level feature values, a subset of high-level feature values associated with the set of uncertain pixels,

wherein redetermining whether the set of uncertain pixels correspond to the one or more objects comprises redetermining whether the set of uncertain pixels correspond to the one or more objects based on the subset of low-level feature values and the subset of high-level feature values.

5. The non-transitory computer-readable medium of claim 3 , wherein the instructions, when executed by the at least one processor, cause the computing device to utilize the segmentation refinement neural network to generate the initial segmentation mask for the digital visual media item by utilizing the segmentation refinement neural network to generate the initial segmentation mask based on the final feature map.

6. The non-transitory computer-readable medium of claim 1 , wherein the instructions, when executed by the at least one processor, cause the computing device to utilize the segmentation refinement neural network to generate the initial segmentation mask for the digital visual media item by utilizing a backbone neural network component of the segmentation refinement neural network to generate the initial segmentation mask, the backbone neural network component comprising a convolutional neural network.

7. The non-transitory computer-readable medium of claim 1 , further comprising instructions that, when executed by the at least one processor, cause the computing device to utilize the segmentation refinement neural network to:

determine, based on the refined segmentation mask, additional uncertain pixels that correspond to a subset of the uncertain pixels; and

generate an additional refined segmentation mask for the digital visual media item by redetermining whether a set of additional uncertain pixels correspond to the one or more objects.

8. The non-transitory computer-readable medium of claim 1 , wherein:

the digital visual media item comprises a digital video feed; and

the instructions, when executed by the at least one processor, cause the computing device to utilize the segmentation refinement neural network to generate the initial segmentation mask for the digital visual media item by utilizing the segmentation refinement neural network to generate the initial segmentation mask for a video frame of the digital video feed.

9. The non-transitory computer-readable medium of claim 1 , further comprising instructions that, when executed by the at least one processor, cause the computing device to modify the digital visual media item based on the refined segmentation mask.

10. A system comprising:

at least one memory device comprising a segmentation refinement neural network; and

at least one server device configured to cause the system to:

receive a digital visual media item comprising a plurality of pixels and depicting one or more objects; and

utilize the segmentation refinement neural network to:

generate a set of feature maps corresponding to the digital visual media item comprising a set of initial feature maps and a final feature map;

generate an initial segmentation mask by determining whether the plurality of pixels correspond to the one or more objects based on the final feature map;

determine, based on the initial segmentation mask, uncertain pixels, the uncertain pixels having an associated uncertainty that the uncertain pixels correspond to the one or more objects or do not correspond to the one or more objects;

extract, from the set of initial feature maps, feature values associated with the uncertain pixels; and

generate a refined segmentation mask for the digital visual media item by redetermining whether a set of uncertain pixels correspond to the one or more objects based on the feature values associated with the uncertain pixels.

11. The system of claim 10 , wherein the at least one server device is configured to cause the system to:

utilize the segmentation refinement neural network to generate the initial segmentation mask based on the final feature map by utilizing a convolutional neural network to generate the initial segmentation mask; and

utilize the segmentation refinement neural network to generate the refined segmentation mask for the digital visual media item by utilizing a multi-layer perceptron renderer to generate the refined segmentation mask.

12. The system of claim 10 , wherein:

the digital visual media item is associated with a first resolution; and

the at least one server device is configured to:

utilize the segmentation refinement neural network to generate the initial segmentation mask by utilizing the segmentation refinement neural network to generate the initial segmentation mask having a second resolution that is lower than the first resolution; and

utilize the segmentation refinement neural network to generate the refined segmentation mask by utilizing the segmentation refinement neural network to generate the refined segmentation mask having a third resolution that is higher than the second resolution.

13. The system of claim 10 , wherein:

the at least one server device is further configured to cause the system to utilize the segmentation refinement neural network to:

extract, from the final feature map, additional feature values associated with the uncertain pixels; and

extract, from the final feature map, feature values associated with certain pixels having an associated certainty that the certain pixels correspond to the one or more objects or do not correspond to the one or more objects; and

redetermining whether the set of uncertain pixels correspond to the one or more objects comprises:

redetermining whether the set of uncertain pixels correspond to the one or more objects based on the feature values associated with the uncertain pixels and the additional feature values associated with the uncertain pixels; and

redetermining whether the certain pixels correspond to the one or more objects based on the feature values associated with the certain pixels.

14. The system of claim 10 , wherein the at least one server device is configured to utilize the segmentation refinement neural network to determine the uncertain pixels by generating, based on the final feature map and the initial segmentation mask, an uncertainty map that provides uncertainty scores for pixels of the initial segmentation mask utilizing one or more neural network layers having learned network weights.

15. The system of claim 14 , wherein the at least one server device is further configured to identify the set of uncertain pixels from the uncertain pixels based on a ranking of the uncertainty scores provided by the uncertainty map.

16. The system of claim 10 , wherein the at least one server device is configured to cause the system to utilize the segmentation refinement neural network to generate the refined segmentation mask for the digital visual media item by utilizing the segmentation refinement neural network to generate a sky mask corresponding to the digital visual media item or a salient mask corresponding to the digital visual media item.

17. The system of claim 10 , wherein:

the digital visual media item comprises a digital video feed;

the at least one server device is configured to cause the system to utilize the segmentation refinement neural network to generate the refined segmentation mask for the digital visual media item by utilizing the segmentation refinement neural network to generate the refined segmentation mask for a first video frame of the digital video feed; and

the at least one server device is further configured to cause the system to utilize the segmentation refinement neural network to generate an additional refined segmentation mask for a second video frame of the digital video feed based on the refined segmentation mask.

18. In a digital medium environment for editing digital visual media, a computer-implemented method for generating segmentation masks comprising:

receiving, from a computing device, a digital visual media item comprising a plurality of pixels, the digital visual media item depicting one or more objects; and

utilizing a segmentation refinement neural network to:

generate an initial segmentation mask corresponding to the digital visual media item by determining whether the plurality of pixels correspond to the one or more objects by generating probabilities that the plurality of pixels correspond to the one or more objects;

determine, based on the initial segmentation mask, uncertain pixels, the uncertain pixels having an associated uncertainty that the uncertain pixels correspond to the one or more objects or do not correspond to the one or more objects; and

generate a refined segmentation mask for the digital visual media item by redetermining whether a set of uncertain pixels correspond to the one or more objects by generating updated probabilities that the plurality of pixels correspond to the one or more objects.

19. The computer-implemented method of claim 18 , wherein:

the computing device comprises a mobile device; and

generating the initial segmentation mask corresponding to the digital visual media item comprises generating the initial segmentation mask at the mobile device.

20. The computer-implemented method of claim 18 , further comprising modifying the digital visual media item utilizing the refined segmentation mask.