Depth map generation

摘要

Depth maps are generated from two or more of images captured with a conventional digital camera from the same viewpoint using different configuration settings, which may be arbitrarily selected for each image. The configuration settings may include aperture and focus settings and/or other configuration settings capable of introducing blur into an image. The depth of a selected image patch is evaluated over a set of discrete depth hypotheses using a depth likelihood function modeled to analyze corresponding images patches convolved with blur kernels using a flat prior in the frequency domain. In this way, the depth likelihood function may be evaluated without first reconstructing an all-in-focus image. Blur kernels used in the depth likelihood function and are identified from a mapping of depths and configuration settings to the blur kernels. This mapping is determined from calibration data for the digital camera used to capture the two or more images.

主权项

1. A computer-implemented method comprising:
receiving, by a computer system, a first image and a second image that both depict a scene; identifying, by the computer system, first image patches from the first image and second image patches from the second image; iteratively evaluating, by the computer system, each of the first image patches over multiple depth hypotheses, wherein the iterative evaluation comprises, for each depth hypothesis and each of the first image patches:
identifying a first blur kernel that is dependent on the depth hypothesis and a first configuration setting used to capture the first image,identifying a second blur kernel that is dependent on the depth hypothesis and a second configuration setting used to capture the second image, anddetermining a depth likelihood for the depth hypotheses based on convolving the image patch with the second blur kernel and convolving a respective one of the second image patches with the first blur kernel; determining, for each of the first image patches, a depth estimate for the image patch based on one of the depth hypotheses for which the iterative evaluation generates a maximal depth likelihood; and generating, by the computer system, a depth map using the depth estimate for each image patch.