Depth estimation for autonomous devices

1. A computer-implemented method of depth estimation, the computer-implemented method comprising:

accessing, by a computing system comprising one or more computing devices, a feature representation associated with a pair of stereo images comprising a first plurality of points and a second plurality of points respectively;

determining, by the computing system, sparse disparity estimates associated with disparities between the first plurality of points and the second plurality of points, wherein the sparse disparity estimates are based at least in part on one or more machine-learned models configured to perform operations comprising estimating the disparities based at least in part on one or more comparisons of portions of the first plurality of points to portions of the second plurality of points;

determining, by the computing system, based at least in part on the sparse disparity estimates and the one or more machine-learned models, a plurality of confidence ranges associated with the disparities between the portions of the first plurality of points and the portions of the second plurality of points;

generating, by the computing system, a disparity map for the pair of stereo images based at least in part on using the plurality of confidence ranges and the one or more machine-learned models to prune the disparities that are outside of the plurality of confidence ranges; and

generating, by the computing system, based at least in part on the disparity map, one or more three-dimensional depth estimates associated with the pair of stereo images.

2. The computer-implemented method of claim 1 , wherein the determining, by the computing system, sparse disparity estimates associated with disparities between the first plurality of points and the second plurality of points of the pair of stereo images comprises:

determining, by the computing system, based at least in part on a random sampling of the first plurality of points, the portions of the first plurality of points to compare against the portions of the second plurality of points, wherein locations of the portions of the second plurality of points correspond to locations of the portions of the first plurality of points;

determining, by the computing system, a plurality of disparities between the portions of the first plurality of points and the portions of the second plurality of points based at least in part on the one or more comparisons of the portions of the first plurality of points to the portions of the second plurality of points; and

determining, by the computing system, the sparse disparity estimates based at least in part on matching scores associated with the plurality of disparities between the portions of the first plurality of points and the portions of the second plurality of points.

3. The computer-implemented method of claim 1 , wherein the random sampling of the first plurality of points comprises a random sampling of a uniform distribution of the first plurality of points.

4. The computer-implemented method of claim 1 , wherein at least one of the one or more machine-learned models comprises a plurality of layers configured to perform the operations comprising the one or more comparisons of the portions of the first plurality of points and the portions of the second plurality of points, and wherein the plurality of layers comprises a sampling layer, a propagation layer, and an evaluation layer.

5. The computer-implemented method of claim 1 , wherein the one or more comparisons of the portions of the first plurality of points to the portions of the second plurality of points comprise use of a one-hot filter pattern on the portions of the first plurality of points and the portions of the second plurality of points.

6. The computer-implemented method of claim 1 , wherein the sparse disparity estimates are determined over a plurality of iterations.

7. The computer-implemented method of claim 6 , wherein each of the plurality of iterations comprises one or more disparity operations associated with determination of the portions of the first plurality of points that have minimal disparity with respect to the portions of the second plurality of points.

8. The computer-implemented method of claim 1 , wherein the one or more machine-learned models are trained based at least in part on minimization of a loss associated with one or more differences between a training disparity map and a ground-truth disparity map, wherein the training disparity map is generated based at least in part on training data and the one or more machine-learned models, and wherein the training data comprises a plurality of training feature representations respectively associated with a plurality of pairs of training stereo images.

9. The computer-implemented method of claim 8 , wherein the loss is based at least in part on a loss function associated with a cost aggregated disparity loss or a refined disparity loss.

10. The computer-implemented method of claim 8 , wherein the loss is positively correlated with one or more disparities between the training disparity map and the ground-truth disparity map.

11. The computer-implemented method of claim 1 , wherein the determining, by the computing system, based at least in part on the sparse disparity estimates and the one or more machine-learned models, a plurality of confidence ranges associated with the disparities between the portions of the first plurality of points and the second plurality of points comprises:

generating, by the computing system, a warped second plurality of points based at least in part on the sparse disparity estimates and the second plurality of points; and

determining, by the computing system, the plurality of confidence ranges based at least in part on the one or more machine-learned models comprising a machine-learned confidence prediction model configured to determine the plurality of confidence ranges based at least in part on an input comprising the first plurality of points, the warped plurality of points, and the sparse disparity estimates.

12. The computer-implemented method of claim 1 , wherein the generating, by the computing system, a disparity map for the pair of stereo images based at least in part on using the plurality of confidence ranges and the one or more machine-learned models to prune the disparities that are outside of the plurality of confidence ranges comprises:

determining, by the computing system, one or more refined disparity predictions based at least in part on the sparse disparity estimates, the first plurality of points, and the one or more machine-learned models comprising a machine-learned refinement model configured to analyze low-level features of the feature representation; and

generating, by the computing system, the disparity map based at least in part on the one or more refined disparity predictions.

13. The computer-implemented method of claim 1 , wherein the pair of stereo images is rectified and comprises overlapping fields of view of an environment.

14. The computer-implemented method of claim 1 , wherein the first plurality of points and the second plurality of points correspond to a first plurality of pixels and a second plurality of pixels respectively.

15. A computing system comprising:

one or more processors;

a memory comprising one or more tangible non-transitory computer-readable media, the memory storing computer-readable instructions that when executed by the one or more processors cause the computing system to perform operations comprising:

accessing a feature representation associated with a pair of stereo images comprising a first plurality of points and a second plurality of points respectively;

determining sparse disparity estimates associated with disparities between the first plurality of points and the second plurality of points, wherein the sparse disparity estimates are based at least in part on one or more machine-learned models configured to perform operations comprising estimating the disparities based at least in part on one or more comparisons of portions of the first plurality of points to portions of the second plurality of points;

determining, based at least in part on the sparse disparity estimates and the one or more machine-learned models, a plurality of confidence ranges associated with the disparities between the portions of the first plurality of points and the portions of the second plurality of points;

generating a disparity map for the pair of stereo images based at least in part on using the plurality of confidence ranges and the one or more machine-learned models to prune the disparities that are outside of the plurality of confidence ranges; and

generating, based at least in part on the disparity map, one or more three-dimensional depth estimates associated with the pair of stereo images.

16. The computing system of claim 15 , wherein the one or more machine-learned models comprise at least one end-to-end differentiable machine-learned model.

17. The computing system of claim 15 , wherein each of the plurality of confidence ranges comprises confidence values associated with one or more disparities between the portions of the first plurality of points and the portions of the second plurality of points.

18. An autonomous vehicle comprising:

one or more processors;

a memory comprising one or more tangible non-transitory computer-readable media, the memory storing computer-readable instructions that when executed by the one or more processors cause a computing system to perform operations comprising:

accessing a feature representation associated with a pair of stereo images comprising a first plurality of points and a second plurality of points respectively;

determining sparse disparity estimates associated with disparities between the first plurality of points and the second plurality of points, wherein the sparse disparity estimates are based at least in part on one or more machine-learned models configured to perform operations comprising estimating the disparities based at least in part on one or more comparisons of portions of the first plurality of points to portions of the second plurality of points;

determining, based at least in part on the sparse disparity estimates and the one or more machine-learned models, a plurality of confidence ranges associated with the disparities between the portions of the first plurality of points and the portions of the second plurality of points;

generating a disparity map for the pair of stereo images based at least in part on using the plurality of confidence ranges and the one or more machine-learned models to prune the disparities that are outside of the plurality of confidence ranges; and

generating, based at least in part on the disparity map, one or more three-dimensional depth estimates associated with the pair of stereo images.

19. The autonomous vehicle of claim 18 , wherein the one or more three-dimensional depth estimates are used as part of an input to a perception system of the autonomous vehicle.

20. The autonomous vehicle of claim 18 , further comprising:

controlling one or more vehicle systems of the autonomous vehicle based at least in part on the one or more three-dimensional depth estimates.