Systems and methods for generating depth maps and corresponding confidence maps indicating depth estimation reliability

• US 9,123,117 B2
• Filed: 10/28/2014
• Issued: 09/01/2015
• Est. Priority Date: 08/21/2012
• Status: Active Grant

• Current Assignee: FotoNation Limited (Adeia Inc.)
• Original Assignee: Pelican Imaging Corporation
• Inventors: Ciurea, Florian, Venkataraman, Kartik, Molina, Gabriel, Lelescu, Dan
• Primary Examiner(s): Tung, Kee M
• Assistant Examiner(s): Du, Haixia
• Application Number:
  US14/526,364
• Publication Number:
  US 20150049915A1
• Time in Patent Office: 308 Days
• US Class Current: 1/1
• CPC Class Codes: G02B 27/0075 with means for altering, e.g. increasing, the depth of field or depth of focus
• CPC Class Codes: G06T 15/20 Perspective computation
• CPC Class Codes: G06T 2200/21 involving computational photography
• CPC Class Codes: G06T 2207/10012 Stereo images
• CPC Class Codes: G06T 2207/10024 Color image
• CPC Class Codes: G06T 2207/10052 Images from lightfield camera
• CPC Class Codes: G06T 7/557 from light fields, e.g. from plenoptic cameras
• CPC Class Codes: G06T 7/593 from stereo images
• CPC Class Codes: G06T 7/85 Stereo camera calibration
• CPC Class Codes: H04N 13/128 Adjusting depth or disparity
• CPC Class Codes: H04N 13/232 using fly-eye lenses, e.g. arrangements of circular lenses
• CPC Class Codes: H04N 13/243 using three or more 2D image sensors
• CPC Class Codes: H04N 2013/0081 Depth or disparity estimation from stereoscopic image signals
• CPC Class Codes: H04N 2013/0088 Synthesising a monoscopic image signal from stereoscopic images, e.g. synthesising a panoramic or high resolution monoscopic image
• CPC Class Codes: H04N 23/16 Optical arrangements associated therewith, e.g. for beam-splitting or for colour correction

Abstract

Systems in accordance with embodiments of the invention can perform parallax detection and correction in images captured using array cameras. Due to the different viewpoints of the cameras, parallax results in variations in the position of objects within the captured images of the scene. Methods in accordance with embodiments of the invention provide an accurate account of the pixel disparity due to parallax between the different cameras in the array, so that appropriate scene-dependent geometric shifts can be applied to the pixels of the captured images when performing super-resolution processing. In a number of embodiments, generating depth estimates considers the similarity of pixels in multiple spectral channels. In certain embodiments, generating depth estimates involves generating a confidence map indicating the reliability of depth estimates.

Claims

1. A method of estimating distances to objects within a scene based upon a set of images captured from different viewpoints using a processor configured by an image processing application, the method comprising:

selecting the viewpoint of an image from the set of images captured from different viewpoints as a reference viewpoint;

normalizing the set of images to increase the similarity of corresponding pixels within the set of images;

determining depth estimates for pixel locations in an image from the reference viewpoint using at least a subset of the set of images, wherein generating a depth estimate for a given pixel location in the image from the reference viewpoint comprises;

identifying pixels in the at least a subset of the set of images that correspond to the given pixel location in the image from the reference viewpoint based upon expected disparity at a plurality of depths;

comparing the similarity of the corresponding pixels identified at each of the plurality of depths; and

selecting the depth from the plurality of depths at which the identified corresponding pixels have the highest degree of similarity as a depth estimate for the given pixel location in the image from the reference viewpoint; and

generating confidence metrics for the depth estimates for pixel locations in the image from the reference viewpoint, where a confidence metric for a given depth estimate for a given pixel location in the image from the reference viewpoint encodes at least one confidence factor indicating the reliability of the given depth estimate including a confidence factor determined using a measure of the signal to noise ratio (SNR) in a region surrounding the given pixel of the image from the set of images captured from the reference viewpoint.

2. The method of claim 1, wherein the confidence metric encodes at least one binary confidence factor.

3. The method of claim 1, wherein the confidence metric encodes at least one confidence factor represented as a range of degrees of confidence.

4. The method of claim 1, wherein the confidence metric encodes a plurality of confidence factors.

5. The method of claim 4, wherein generating a depth estimate for a given pixel location in the image from the reference viewpoint further comprises:

determining an initial depth estimate for the pixel location in an image from the reference viewpoint based upon the disparity at which the identified corresponding pixels have the highest degree of similarity;

comparing the similarity of the identified corresponding pixels to detect mismatched pixels;

when an initial depth estimate does not result in the detection of a mismatch between the identified corresponding pixels, selecting the initial depth estimate as the current depth estimate for the pixel location in the image from the reference viewpoint; and

when an initial depth estimate results in the detection of a mismatch between the identified corresponding pixels, updating the depth estimate for the pixel location in the image from the reference viewpoint by;

determining a set of candidate depth estimates using a plurality of competing subsets of the set of images based upon the disparities at which corresponding pixels have the highest degree of similarity in each of a plurality of competing subsets of images; and

selecting the candidate depth of the subset having the corresponding pixels with the highest degree of similarity as the updated depth estimate for the pixel location in the image from the reference viewpoint.

6. The method of claim 5, wherein the confidence metric for the depth estimate for a given pixel location in the image from the reference viewpoint comprises at least one confidence factor determined based upon the number of corresponding pixels used to generate the depth estimate.

7. The method of claim 5, wherein the confidence metric for the depth estimate for a given pixel location in the image from the reference viewpoint comprises at least one confidence factor generated by comparing the highest degree of similarity of corresponding pixels from different competing subsets of the set of images.

8. The method of claim 5, wherein the confidence metric encodes at least one confidence factor determined by comparing the similarity of the pixels in the set of images that were used to generate the updated depth estimate for a given pixel location in the image from the reference viewpoint.

9. The method of claim 8, wherein:

a cost function is utilized to generate a cost metric indicating the similarity of corresponding pixels; and

comparing the similarity of the pixels in the set of images that were used to generate the depth estimate for a given pixel location in the image from the reference viewpoint further comprises;

applying a threshold to a cost metric of the pixels in the set of images that were used to generate the updated depth estimate for a given pixel location in the image from the reference viewpoint; and

when the cost metric exceeds the threshold, assigning a confidence metric that indicates that the updated depth estimate for the given pixel location in the image from the reference viewpoint was generated using at least one pixel in the set of images that is a problem pixel.

10. The method of claim 9, wherein the threshold is modified based upon at least one of:

a mean intensity of a region surrounding the given pixel location in the image from the reference viewpoint; and

noise statistics for at least one sensor used to capture the set of images.

11. The method of claim 10, wherein the mean intensity of a region surrounding the given pixel location in the image from the reference viewpoint is calculated using a spatial box averaging filter centered around the given pixel.

12. The method of claim 11, wherein:

the set of images are captured within multiple color channels including at least Red, Green and Blue color channels;

selecting a reference viewpoint relative to the viewpoints of the set of images captured from different viewpoints comprises selecting one of the images in the Green color channel as the reference viewpoint; and

the mean intensity is used to determine the noise statistics for the Green channel using a table that relates a particular mean at a particular exposure and gain to a desired threshold.

13. The method of claim 5, wherein:

a cost function is utilized to generate a cost metric indicating the similarity of corresponding pixels;

a confidence metric based upon general mismatch is obtained using the following formula;

Confidence(x,y)=F(Cost_min(x,y),Cost^d(x,y),I(x,y)^cam,Sensor,Camera Intrinsics)

where Cost_min(x, y) is the cost metric of the pixels in the set of images that were used to generate the depth estimate;

Cost^d(x, y) is the distribution of the cost metrics at pixel location (x, y) for different depths;

I(x, y)^cam image data captured by any camera can be utilized to augment the confidence;

Sensor is the sensor prior, which includes known properties of the sensor; and

Camera intrinsics is the camera intrinsic, which specifies elements intrinsic to the camera and camera array that can impact confidence.

14. The method of claim 5, wherein:

a cost function is utilized to generate a cost metric indicating the similarity of corresponding pixels; and

a confidence metric based upon general mismatch is obtained using the following formula;

15. The method of claim 4, wherein the confidence metric for the depth estimate for a given pixel location in the image from the reference viewpoint further comprises at least one confidence factor selected from the group consisting of:

an indication that the given pixel is within a textureless region within an image;

a number of corresponding pixels used to generate the depth estimate;

an indication of a number of depths searched to generate the depth estimate;

an indication that the given pixel is adjacent a high contrast edge;

an indication that the given pixel is adjacent a high contrast boundary;

an indication that the given pixel lies on a gradient edge;

an indication that corresponding pixels to the given pixel are mismatched;

an indication that corresponding pixels to the given pixel are occluded;

an indication that depth estimates generated using different reference cameras exceed a threshold for the given pixel; and

an indication that depth estimates generated using different subsets of cameras exceed a threshold for the given pixel;

an indication as to whether the depth of the given threshold exceeds a threshold;

an indication that the given pixel is defective; and

an indication that corresponding pixels to the given pixel are defective.

16. The method of claim 1, wherein generating confidence metrics for the depth estimates for pixel locations in the image from the reference viewpoint comprises determining at least one sensor gain used to capture at least one of the set of images and adjusting at least one confidence factor based upon a sensor gain.

17. The method of claim 1, generating confidence metrics for the depth estimates for pixel locations in the image from the reference viewpoint comprises determining at least one exposure time used to capture at least one of the set of images and adjusting at least one confidence factor based upon an exposure time.

18. The method of claim 1, further comprising:

outputting a depth map containing the depth estimates for pixel locations in the image from the reference viewpoint; and

outputting a confidence map containing confidence metrics for the updated depth estimates contained within the depth map.

19. The method of claim 18, further comprising filtering the depth map based upon the confidence map.

20. The method of claim 1, wherein generating confidence metrics for the depth estimates for pixel locations in the image from the reference viewpoint further comprises encoding a confidence factor for the depth estimate of a specific pixel location in the image from the reference viewpoint by:

calculating local statistics of a region of interest around the specific pixel location in the image from the reference viewpoint; and

comparing the calculated local statistics to local statistics of a similar region in at least one other image from the set of images.

21. The method of claim 20, wherein the confidence factor for the depth estimate of a specific pixel location in the image from the reference viewpoint is a function of the mean and variance of the regions across images.

22. A method of synthesizing a higher resolution image from a set of lower resolution images captured from different viewpoints, the method comprising:

estimating distances to objects within a scene from a set of images captured from different viewpoints using a processor directed by an image processing application to;

select the viewpoint of an image from the set of images captured from different viewpoints as a reference viewpoint;

normalize the set of images to increase the similarity of corresponding pixels within the set of images;

determine depth estimates for pixel locations in an image from the reference viewpoint using at least a subset of the set of images, wherein generating a depth estimate for a given pixel location in the image from the reference viewpoint comprises;

identifying pixels in the at least a subset of the set of images that correspond to the given pixel location in the image from the reference viewpoint based upon expected disparity at a plurality of depths;

comparing the similarity of the corresponding pixels identified at each of the plurality of depths; and

selecting the depth from the plurality of depths at which the identified corresponding pixels have the highest degree of similarity as a depth estimate for the given pixel location in the image from the reference viewpoint; and

generating confidence metrics for the depth estimates for pixel locations in the image from the reference viewpoint using a processor directed by an image processing application, where a confidence metric for a given depth estimate for a pixel location in the image from the reference viewpoint encodes at least one confidence factor indicating the reliability of the given depth estimate;

determining the visibility of the pixels in the set of images from the reference viewpoint using a processor directed by an image processing application; and

fusing pixels from the set of images using the processor directed by the image processing application based upon the depth estimates that are indicated as being reliable by the confidence metrics to create a fused image having a resolution that is greater than the resolutions of the images in the set of images by;

identifying the pixels from the set of images that are visible in an image from the reference viewpoint and that have reliable depth estimates using the visibility information and the confidence metrics; and

applying scene dependent geometric shifts to the pixels from the set of images that are visible in an image from the reference viewpoint and that have reliable depth estimates to shift the pixels into the reference viewpoint, where the scene dependent geometric shifts are determined using the depth estimates; and

fusing the shifted pixels from the set of images to create a fused image from the reference viewpoint having a resolution that is greater than the resolutions of the images in the set of images.

23. The method of claim 22, further comprising synthesizing an image from the reference viewpoint using the processor directed by the image processing application to perform a super-resolution process based upon the fused image from the reference viewpoint, the set of images captured from different viewpoints, the depth estimates, the visibility information, and the confidence metrics.

24. The method of claim 22, wherein a confidence metric for a given depth estimate for a pixel location in the image from the reference viewpoint encodes at least one confidence factor determined using a measure of the signal to noise ratio (SNR) in a region surrounding a given pixel of the image from the set of images captured from the reference viewpoint.

25. The method of claim 22, wherein generating confidence metrics for the depth estimates for pixel locations in the image from the reference viewpoint further comprises encoding a confidence factor for the depth estimate of a specific pixel location in the image from the reference viewpoint by:

calculating local statistics of a region of interest around the specific pixel location in the image from the reference viewpoint; and

comparing the calculated local statistics to the local statistics of a similar region in at least one other image from the set of images.

26. The method of claim 25, wherein the confidence factor for the depth estimate of a specific pixel location in the image from the reference viewpoint is a function of the mean and variance of the regions across images.

27. The method of claim 22, wherein the confidence metric encodes a plurality of confidence factors.

28. The method of claim 27, wherein the confidence metric for the depth estimate for a given pixel location in the image from the reference viewpoint further comprises at least one confidence factor selected from the group consisting of:

an indication that the given pixel is within a textureless region within an image;

a number of corresponding pixels used to generate the depth estimate;

an indication of a number of depths searched to generate the depth estimate;

an indication that the given pixel is adjacent a high contrast edge;

an indication that the given pixel is adjacent a high contrast boundary;

an indication that the given pixel lies on a gradient edge;

an indication that corresponding pixels to the given pixel are mismatched;

an indication that corresponding pixels to the given pixel are occluded;

an indication that depth estimates generated using different reference cameras exceed a threshold for the given pixel; and

an indication that depth estimates generated using different subsets of cameras exceed a threshold for the given pixel;

an indication as to whether the depth of the given threshold exceeds a threshold;

an indication that the given pixel is defective; and

an indication that corresponding pixels to the given pixel are defective.

29. An image processing system, comprising:

a processor;

memory containing a set of images captured from different viewpoints and an image processing application;

wherein the image processing application directs the processor to;

select a viewpoint of an image from the set of images captured from different viewpoints as a reference viewpoint;

normalize the set of images to increase the similarity of corresponding pixels within the set of images;

determine depth estimates for pixel locations in an image from the reference viewpoint using at least a subset of the set of images, wherein generating a depth estimate for a given pixel location in the image from the reference viewpoint comprises;

identify pixels in the at least a subset of the set of images that correspond to the given pixel location in the image from the reference viewpoint based upon expected disparity at a plurality of depths;

compare the similarity of the corresponding pixels identified at each of the plurality of depths; and

select the depth from the plurality of depths at which the identified corresponding pixels have the highest degree of similarity as a depth estimate for the given pixel location in the image from the reference viewpoint; and

generate confidence metrics for the depth estimates for pixel locations in the image from the reference viewpoint, where a confidence metric for a given depth estimate for a given pixel location in the image from the reference viewpoint encodes at least one confidence factor indicating the reliability of the given depth estimate including a confidence factor determined using a measure of the signal to noise ratio (SNR) in a region surrounding the given pixel from the reference viewpoint.