Systems and methods for hybrid depth regularization

1. A depth sensing system, comprising:

a plurality of cameras;

a processor;

a memory containing an image processing application;

wherein the image processing application directs the processor to:

obtain image data for a plurality of images from multiple viewpoints using the plurality of cameras, wherein the image data for the plurality of images comprises a reference image and at least one alternate view image;

generate a raw depth map containing depth estimates for pixels within the reference image using the image data for the reference image and the image data for the at least one alternate view image using a first depth estimation process, and a confidence map describing reliability of depth estimates contained within the raw depth map; and

generate a regularized depth map by:

computing a secondary depth map containing depth estimates for pixels within the reference image using a second depth estimation process, wherein the second depth estimation process calculates depth using the image data for the reference image and the image data for the at least one alternate view image, by using a different estimation technique than that of the first depth estimation process; and

computing a composite depth map by selecting depth estimates from the raw depth map and the secondary depth map, where a depth estimate for a pixel in the reference image is selected from the raw depth map when the depth estimate is indicated as being reliable by the confidence map.

2. The depth sensing system of claim 1 , wherein the composite depth map is a first composite depth map and the image processing application further directs the processor to generate a regularized depth map by:

computing a second composite depth map by:

applying a smoothing filter to depth estimates in the first composite depth map;

computing an edge map using the filtered first composite depth map, where the edge map indicates pixels within the reference image that form part of an edge; and

computing a second composite depth map by selecting depth estimates from the raw depth map and the filtered first composite depth map, where a depth estimate for a pixel in the reference image is selected from the raw depth map when the edge map indicates that the depth estimate is with respect to a pixel from the reference image that forms part of an edge.

3. The depth sensing system of claim 2 , wherein the image processing application further directs the processor to apply an edge preserving filter to at least one region of the second composite depth map indicated as containing a depth discontinuity by the edge map.

4. The depth sensing system of claim 2 , wherein the image processing application further directs the processor to compute the second composite depth map by:

applying edge detection to the filtered first composite depth map to result in an edge-detected depth map; and

dilating the edge-detected depth map to result in the edge map.

5. The depth sensing system of claim 1 , wherein the image processing application further directs the processor to generate a raw depth map containing depth estimates for pixels within the reference image using the image data for the reference image and the image data for the at least one alternate view image, and a confidence map describing reliability of depth estimates contained within the depth map, by measuring parallax observable between the reference image and the at least one alternate view image.

6. The depth sensing system of claim 5 , wherein the image processing application further directs the processor to estimate depth based upon measured parallax observable between the reference image and the at least one alternate view image by:

measuring parallax by comparing the similarity of a pixel in the reference image to pixels in the at least one alternate view image determined based upon a plurality of depth samples using a cost function; and

estimating depth for the pixel in the reference image by identifying the sampled depth at which the cost function for a pixel in the reference image indicates the strongest match as being the estimated depth of the pixel.

7. The depth sensing system of claim 6 , wherein the image processing application further directs the processor to:

construct a cost volume in memory using costs determined using the cost function at each sampled depth; and

compute the secondary depth map based on the cost volume.

8. The depth sensing system of claim 7 , wherein the image processing application further directs the processor to compute the secondary depth map using a second depth estimation process comprising:

downsampling the reference image and the cost volume; and

estimating depths to enforce smooth depth transitions within the secondary depth map in textureless regions of the reference image using the cost volume.

9. The depth sensing system of claim 1 , wherein the image processing application further directs the processor to compute the composite depth map by determining that a depth estimate for a pixel from the raw depth map is reliable when a confidence value for the pixel from the confidence map is above a threshold.

10. The depth sensing system of claim 1 , wherein the image processing application further directs the processor to:

identify at least one corresponding region within the reference image and the at least one alternate view image using at least one criterion;

generate binary images with respect to the reference image and the at least one alternate view image based upon the identified at least one corresponding region;

subtract shifted versions of the reference image and the at least one alternate view image when the shifts correspond to a plurality of depth samples to produce a shifted and subtracted image for each of the plurality of depth samples;

identify boundaries within the shifted and subtracted images;

identify areas of the boundaries within the shifted and subtracted images; and

determine depth estimates for pixels at a boundary of an identified region in the reference image based upon the depth sample at which the area of the boundary in the shifted and subtracted image is minimized.

11. A depth sensing method, comprising:

obtaining image data for a plurality of images from multiple viewpoints using the plurality of cameras, wherein the image data for the plurality of images comprises a reference image and at least one alternate view image;

generating a raw depth map containing depth estimates for pixels within the reference image using the image data for the reference image and the image data for the at least one alternate view image using a first depth estimation process, and a confidence map describing reliability of depth estimates contained within the raw depth map; and

generating a regularized depth map by:

computing a secondary depth map containing depth estimates for pixels within the reference image using a second depth estimation process, wherein the second depth estimation process calculates depth using the image data for the reference image and the image data for the at least one alternate view image, by using a different estimation technique than that of the first depth estimation process; and

computing a composite depth map by selecting depth estimates from the raw depth map and the secondary depth map, where a depth estimate for a pixel in the reference image is selected from the raw depth map when the depth estimate is indicated as being reliable by the confidence map.

12. The depth sensing method of claim 11 , wherein the composite depth map is a first composite depth map and generating a regularized depth map further comprises:

computing a second composite depth map by:

applying a smoothing filter to depth estimates in the first composite depth map;

computing an edge map using the filtered first composite depth map, where the edge map indicates pixels within the reference image that form part of

an edge; and

computing a second composite depth map by selecting depth estimates from the raw depth map and the filtered first composite depth map, where a depth estimate for a pixel in the reference image is selected from the raw depth map when the edge map indicates that the depth estimate is with respect to a pixel from the reference image that forms part of an edge.

13. The depth sensing method of claim 12 , further comprising applying an edge preserving filter to at least one region of the second composite depth map indicated as containing a depth discontinuity by the edge map.

14. The depth sensing method of claim 12 , wherein computing the second composite depth map further comprises:

applying edge detection to the filtered first composite depth map to result in an edge-detected depth map; and

dilating the edge-detected depth map to result in the edge map.

15. The depth sensing method of claim 11 , wherein generating a raw depth map containing depth estimates for pixels within the reference image using the image data for the reference image and the image data for the at least one alternate view image, and a confidence map describing reliability of depth estimates contained within the depth map, further comprises measuring parallax observable between the reference image and the at least one alternate view image.

16. The depth sensing method of claim 15 , further comprising estimating depth based upon measured parallax observable between the reference image and the at least one alternate view image by:

measuring parallax by comparing the similarity of a pixel in the reference image to pixels in the at least one alternate view image determined based upon a plurality of depth samples using a cost function; and

estimating depth for the pixel in the reference image by identifying the sampled depth at which the cost function for a pixel in the reference image indicates the strongest match as being the estimated depth of the pixel.

17. The depth sensing method of claim 16 , further comprising:

generating cost volume using costs determined using the cost function at each sampled depth; and

computing the secondary depth map based on the cost volume.

18. The depth sensing method of claim 17 , wherein computing the secondary depth map using a second depth estimation process comprising:

downsampling the reference image and the cost volume; and

estimating depths to enforce smooth depth transitions within the secondary depth map in textureless regions of the reference image using the cost volume.

19. The depth sensing method of claim 11 , wherein computing the composite depth map further comprises determining that a depth estimate for a pixel from the raw depth map is reliable when a confidence value for the pixel from the confidence map is above a threshold.

20. The depth sensing method of claim 11 , further comprising:

identifying at least one corresponding region within the reference image and the at least one alternate view image using at least one criterion;

generating binary images with respect to the reference image and the at least one alternate view image based upon the identified at least one corresponding region;

subtracting shifted versions of the reference image and the at least one alternate view image when the shifts correspond to a plurality of depth samples to produce a shifted and subtracted image for each of the plurality of depth samples;

identifying boundaries within the shifted and subtracted images;

identifying areas of the boundaries within the shifted and subtracted images; and

determining depth estimates for pixels at a boundary of an identified region in the reference image based upon the depth sample at which the area of the boundary in the shifted and subtracted image is minimized.