High quality instance segmentation

1. A computer-implemented method comprising:

obtaining an image comprising a plurality of pixels;

generating, using a first portion of a machine-learned segmentation model, an initial distance function indicative of a distance of a respective pixel, of the plurality of pixels, to a predicted nearest boundary of an object depicted in the image;

generating, using a second portion of the machine-learned segmentation model, an object feature representation that comprises a feature embedding for the image;

determining, using an energy function, a final distance function based on the initial distance function and the object feature representation,

wherein the final distance function indicates that the respective pixel is associated with a background of the image or that the respective pixel is associated with the object depicted in the image; and

determining an instance segmentation mask for the image based on the final distance function.

2. The computer-implemented method of claim 1 , wherein generating the initial distance function indicative of the distance of the respective pixel comprises generating a truncated signed distance function initialization that estimates a signed distance for the respective pixel to another pixel that is associated with the predicted nearest boundary of the object depicted in the image.

3. The computer-implemented method of claim 2 , further comprising:

iteratively updating, based on the feature representation, the truncated signed distance function initialization associated with the object.

4. The computer-implemented method of claim 3 , wherein iteratively updating the truncated signed distance function initialization comprises providing the truncated signed distance function initialization and the object feature representation to a recurrent neural network.

5. The computer-implemented method of claim 1 , wherein generating the object feature representation comprises generating the feature embedding based on an intensity of the respective pixel.

6. The computer-implemented method of claim 1 , further comprising:

identifying, based on the instance segmentation mask, a region of pixels included in the plurality of pixels that are associated with the object.

7. The computer-implemented method of claim 1 , further comprising:

determining, using a third portion of the machine-learned segmentation model, a label for the object depicted in the image.

8. The computer-implemented method of claim 7 , wherein the label is indicative of a type of the object.

9. A computing system comprising:

one or more processors; and

one or more non-transitory computer-readable media that store instructions that are executable by one or more processors to cause the one or more processors to perform operations comprising:

obtaining an image comprising a plurality of pixels;

generating, using a first portion of a machine-learned segmentation model, an initial distance function indicative of a distance of a respective pixel, of the plurality of pixels, to a predicted nearest boundary of an object depicted in the image;

generating, using a second portion of the machine-learned segmentation model, an object feature representation that comprises a feature embedding for the image;

determining, using an energy function, a final distance function based on the initial distance function and the object feature representation,

wherein the final distance function indicates that the respective pixel is associated with a background of the image or that the respective pixel is associated with the object depicted in the image; and

determining an instance segmentation mask for the image based on the final distance function.

10. The computing system of claim 9 , wherein the initial distance function comprises a signed distance function initialization.

11. The computing system of claim 10 , wherein the signed distance function initialization estimates a signed distance for the respective pixel to another pixel that is associated with the predicted nearest boundary of the object depicted in the image.

12. The computing system of claim 9 , further comprising:

iteratively updating the initial distance function associated with the object based on the feature representation.

13. The computing system of claim 9 , wherein the first portion of the machine-learned segmentation model comprises a first head configured to generate the initial distance function for the object depicted in the image, and the second portion of the machine-learned segmentation model comprises a second head configured to output the object feature representation comprising the feature embedding.

14. The computing system of claim 13 , wherein the machine-learned segmentation model comprises a third portion, the third portion comprising a third head configured to classify and localize the object.

15. The computing system of claim 9 , wherein the image is obtained through a sensor of an autonomous vehicle and wherein the object is within a surrounding environment of the autonomous vehicle.

16. The computing system of claim 15 , wherein the autonomous vehicle is an autonomous truck.

17. An autonomous vehicle comprising:

one or more processors; and

one or more non-transitory computer-readable media that store instructions that are executable by one or more processors to cause the one or more processors to perform operations comprising:

obtaining sensor data comprising a plurality of pixels;

generating, using a first portion of a machine-learned segmentation model, an initial distance function indicative of a distance of a respective pixel, of the plurality of pixels, to a predicted nearest boundary of an object depicted in the sensor data, wherein the object is within a surrounding environment of the autonomous vehicle;

generating, using a second portion of the machine-learned segmentation model, an object feature representation that comprises a feature embedding for the sensor data;

determining, using an energy function, a final distance function based on the initial distance function and the object feature representation,

wherein the final distance function indicates that the respective pixel is associated with a background of the sensor data or that the respective pixel is associated with the object depicted in the sensor data; and

determining an instance segmentation mask for the sensor data based on the final distance function.

18. The autonomous vehicle of claim 17 , wherein the operations further comprise:

identifying, based on the instance segmentation mask, a type of the object depicted in the image.

19. The autonomous vehicle of claim 17 , wherein the operations further comprise:

determining a motion trajectory for the autonomous vehicle based on the object depicted in the image.

20. The autonomous vehicle of claim 17 , wherein the sensor data comprises an image obtained through a camera located onboard the autonomous vehicle.