Systems and methods for estimating and refining depth maps

1. A method for improving accuracy of depth map information derived from image data descriptive of a scene where pixels of such image data, acquired with one or more image acquisition devices, each have an assigned intensity value, the method comprising:

performing a matching cost optimization by iteratively refining disparities between corresponding pixels in the image data and using optimization results to create a sequence of first disparity values for an initial disparity map for the scene based in part on a superpixel-wise cost function;

performing a guided filter operation on the first disparity values by applying other image data containing structural details that can be transferred to the first disparity values to restore degraded features or replace some of the first disparity values with values more representative of structural features present in the image data descriptive of the scene,

the guided filtering operation performed by applying a series of weighted median filter operations to pixel intensity values in the sequence of first disparity values so that each median filter operation replaces a member in the sequence of first disparity values with a median intensity value, where each median intensity value is based on intensity values in a group of pixels within a window of pixels positioned about said member in the sequence,

each window being of a variable size to include a variable number of pixels positioned about said member in the sequence, where selections of multiple ones of the window sizes are based on a measure of similarity between the first disparity values and said other image data, and wherein the series of weighted median filter operations provides a new sequence of disparity values for a refined disparity map or from which a depth map of improved accuracy can be created.

2. The method of claim 1 where the other image data applied to perform the guided filtering operation comprises a portion of the pixel image data acquired with the one or more image acquisition devices.

3. The method of claim 1 where the depth map information is derived from stereo image data comprising data from first and second RBG reference images descriptive of the scene and said other image data containing structural details applied to perform the guided filtering operation comprises data from one of the RBG reference images.

4. The method of claim 1 where the series of median filter operations are weighted median filter operations and multiple ones of the window sizes are selected based on a similarity index values.

5. The method of claim 1 where multiple ones of the window sizes are selected based on measures of differences and similarities between data in the sequence of first disparity values for the initial disparity map and data in one the RGB reference images.

6. The method of claim 1 where multiple ones of the window sizes are selected based on one or more measures of similarity between two sets of image data.

7. The method of claim 6 where the two sets of image data comprise the first disparity values and said other image data.

8. The method of claim 1 where:

multiple ones of the window sizes are selected based on one or more measures of similarity between the first disparity values and said other image data; and

the one or more measures of similarity are based on at least one feature taken from the group consisting of luminance, contrast and structure.

9. The method of claim 8 where the measures of similarity are expressable as a map providing pixel by pixel measures of similarity between the two sets of image data.

10. The method of claim 8 where the measures of similarity are calculated in accord with

SSIM( D,I )=[ I ( D,I )] α ·[ c ( D,I )] β ·[ s ( D,I )] Y ,

where D and I are, respectively, the sequence of first disparity values for the initial disparity map and data values from the image data descriptive of a scene and acquired with said one or more image acquisition devices,

where l, ƒvc and s are, respectively, luminance, contrast and structural terms and

where α, β and γ are exponents for the luminance, contrast, and structural terms, respectively.

11. A method for performing a filter operation on a sequence of data values which are a subset in a larger plurality of data values derived from acquisitions with one or more devices, the method comprising:

performing a series of filter operations on at least some values in the sequence and replacing each of the multiple ones of the values in the sequence with a new value resulting from the filter operation,

where each new value is based on values in a group containing a number of other data values in the sequence and where the number of data values in each group of values is variable and definable in terms of a selectable dimension in a range of window dimensions, and where the window dimensions vary among multiple groups based on a measure of similarity between values in the sequence and other values in the larger plurality of data values, and

where the larger plurality of data values includes image data descriptive of a scene acquired with one or more image acquisition devices.

12. The method of claim 11 where:

the larger plurality of data values includes first and second RGB images comprising image data descriptive of a scene acquired with multiple image acquisition devices; and

each new value is based on values in a different group containing a number of other data values in the sequence.

13. The method of claim 11 where the sequence of data values comprises disparity map data values for an initial disparity map for the scene created from the image data with a matching cost optimization based in part on a superpixel-wise cost function.

14. The method of claim 11 where the filter operation is a median filter operation.

15. The method of claim 11 where the filter operation is a weighted median filter operation.

16. The method of claim 11 where each new value is a median intensity value based on values in the plurality of data values different from data values in the sequence.

17. The method of claim 11 where each new value is a median intensity value based on RBG image data in the plurality of data values different from data values in the sequence.

18. The method of claim 11 where each group of values corresponds to a different set of pixel image data within a two dimensional window positioned about a pixel position for which a value is replaced by a new value based on the filter operation.

19. A method of defining a series of sizes of adaptive windows, each for use in a filter operation performed on data values for a disparity map, D, corresponding to a scene represented by a reference image I, each filter operation replacing a different data value associated with the disparity map with a new value to improve depth map accuracy for the scene, comprising:

calculating data values for the disparity map associated with the scene;

providing data values for the reference image;

providing a plurality of values, S, collectively corresponding to a similarity map, each value, S, indicative of a level of similarity between a portion of the disparity map data values and a corresponding portion of the reference image data values;

applying the values, S, to determine a series of windows of variable sizes, each window in the series for use in a different filter operation performed on the disparity map data values;

with W designating current window size under evaluation:

defining a smallest window size W min of dimension (2j+1) by (2j+1) where j=j min ;

defining a largest window size W max of dimension (2j+1) by (2j+1) where j=j max ;

for each pixel position P i,j in the disparity map, performing a test with the inequality statement

SSIM min <SSIM med <SSIM max

for each corresponding pixel position S i,j in the similarity map beginning with the smallest window size, where:

SSIM min is the minimum pixel value inside the current window, W

SSIM max is the maximum pixel value inside the current window, W and

SSIM med is the median pixel value determined for the current window, W; and concluding that:

If the inequality statement is true, then the current window size is the chosen window size for performing a weighted median operation about the pixel position P i,j in the disparity map corresponding to the similarity map pixel position S i,j ; and

If the inequality statement in step a) is not true, increasing the size of the window by incrementing j to j+1; and

repeating steps a) and b) until either:

(i) SSIM med is between SSIM min and SSIM max , in which case the current window size is the chosen window size for performing a weighted median operation about the pixel position P i,j in the disparity map corresponding to the similarity map pixel position S i,j , or

(ii) the maximum window size is reached, in which case the maximum window size is the chosen window size for performing the weighted median operation about the pixel position P i,j in the disparity map corresponding to the similarity map pixel position S i,j .

20. A method for improving accuracy of depth map information derived from image data descriptive of a scene, comprising:

performing a matching cost optimization by iteratively refining disparities between corresponding pixels in the image data and using optimization results to create a sequence of initial disparity map data values for an initial disparity map for the scene;

creating a Mutual Feature Map data based on the initial disparity map data values and the image data descriptive of the scene;

applying the Mutual Feature Map data to create a series of weighting functions representing structural details that can be transferred to the first disparity values to restore degraded features or replace some of the first disparity values with values more representative of structural features present in the image data descriptive of the scene, including applying a series of weighted median filter operations in which a weight kernel based on the Mutual Feature Map data is applied to assign a weighting factor to each pixel position in a weighted median filter operation, whereby each median filter operation replaces a member in the sequence of first disparity values with a median intensity value, where each median intensity value is based on intensity values in a group of pixels within a window of pixels positioned about a data value in the sequence.