Confidence map, method for generating the same and method for refining a disparity map

1. A method for generating a confidence map, the method being implemented in a disparity estimation stage comprising a hardware processor configured to:

retrieve a first stereo image and a second stereo image, each having a plurality of pixels;

generate a disparity map from the first stereo image and the second stereo image, the disparity map having a plurality of disparity values;

determine a first confidence value for each disparity value based on a match quality between a pixel or a group of pixels in the first stereo image and a corresponding pixel or a corresponding group of pixels in the second stereo image;

determine a second confidence value for each disparity value based on a consistency of disparity estimates of the pixel or the group of pixels in the first stereo image and the corresponding pixel or the corresponding group of pixels in the second stereo image;

wherein the second confidence value assumes one of at least four discrete values of confidence, the discrete values of confidence being RELIABLE, UNRELIABLE, OCCLUDED and UNDECIDED;

determine a combined confidence value for each disparity value from the first confidence value and the second confidence value; and

store the combined confidence value for each disparity value in a confidence map.

2. The method according to claim 1 , wherein the consistency is determined by performing a pixel-based left right consistency check.

3. The method according to claim 1 , wherein the match quality values are re-normalized.

4. The method according to claim 1 , wherein for determination of the second confidence value, information about a visibility of the respective pixel is taken into account.

5. The method according to claim 1 , wherein the value of confidence

RELIABLE is assigned to pixels having a left-right consistency of zero or invisible pixels having a left-right consistency of one pixel;

UNRELIABLE is assigned to visible pixels having a left-right consistency greater than a threshold value;

OCCLUDED is assigned to invisible pixels having a left-right consistency greater than the threshold value; and

UNDECIDED is assigned to all remaining pixels.

6. The method according to claim 5 , wherein the threshold value is two pixels.

7. The method according to claim 1 , wherein the combined confidence value is determined by further taking into account information about a confidence value of further pixels that are located in a surrounding area of the respective pixel.

8. The method according to claim 5 , wherein a confidence value is determined based on at least one of a distance to the next unreliable/occluded pixel and a distance to the next undecided level.

9. A method for refining a disparity map using a confidence map which is associated to the disparity map, the method being implemented in a disparity map refinement stage comprising a hardware processor configured to:

apply a domain filter kernel on a pixel of interest and a pixel in a surrounding area of the pixel of interest;

apply a range filter kernel on the pixel of interest and the pixel in the surrounding area of the pixel of interest;

determine a product of the result of the domain filter kernel and a result of the range filter kernel;

determine a filtered disparity value for the pixel in the surrounding area of the pixel of interest by multiplying a disparity value from the disparity map for the pixel in the surrounding area of the pixel of interest with the product of the result of the domain filter kernel and the result of the range filter kernel and with a confidence filter kernel for the pixel in the surrounding area of the pixel of interest, wherein the confidence filter kernel is derived from the associated confidence map;

generate a refined disparity value for the pixel of interest by determining a weighted average of the filtered disparity value of the pixels in the surrounding area of the pixel of interest; and

store the refined disparity value for each pixel of interest in a refined disparity map.

10. The method according to claim 9 , wherein applying the domain filter kernel comprises determining a spatial distance between the pixel in the surrounding area and the pixel of interest, and wherein applying the range filter kernel comprises determining a color or intensity difference between the pixel in the surrounding area and the pixel of interest.