Estimation of image relations from point correspondences between images
AssigneeHer Majesty the Queen in right of Canada, as represented by the Minister of Industry, through the Communications Research Centre Canada
InventorsLiang Zhang, Demin Wang, Houman Rastgar, André Vincent
Life of this patent
Estimating the term…
Abstract
The invention relates to a method for estimating an image relation model for two or more related images based on a plurality of point correspondences. The method includes using a weighted random sampling algorithm to repeatedly draw subsets of point correspondences for generating model estimates, evaluating said model estimates for the full plurality of the point correspondences, and updating weights associated with each of the point correspondences in the random sampling algorithm based on information obtained in said evaluating.
Claims20
In a system comprising a processor and a memory for storing a plurality of point correspondences (PCs) between two images of a same scene or object obtained from one or more cameras, a method for determining an output image relation model relating the two images, comprising:
- a) using the processor to perform a series of PC sampling steps, each PC sampling step comprising:
- i) selecting a PC subset from the plurality of PCs utilizing a weighted random or pseudo-random selection algorithm wherein individual PCs are selected in dependence upon current values of weight parameters assigned thereto,
- ii) computing an image relation model estimate based on the PC subset, and
- iii) obtaining information related to an accuracy of the image relation model estimate for the plurality of PCs;
- b) obtaining, with the processor, the output image relation model using the information related to the accuracy of the computed image relation model estimates;
- and,
- c) saving the output image relation model in a computer-readable memory for generating new images;
- wherein the current values of the weight parameters assigned to the PCs remain constant in a first K PC sampling steps, so that same current values of the weight parameters are used for same PCs in selecting the PC subsets in the first K PC sampling steps, wherein K is an integer greater than 1,
- and wherein in at least one of the PC sampling steps following the first K PC sampling steps, the current values of the weight parameters are updated by the processor using the image relation model estimate computed in said at least one of the PC sampling steps.
In a system comprising a processor, a memory for storing images, said images comprising two images of a same scene or object obtained from one or more cameras, and further comprising a memory for storing a plurality of point correspondences (PCs) between the two images, a method for determining an output image relation model for the two images, comprising:
- a) using the processor to perform a series of PC sampling steps, each PC sampling step comprising:
- i) selecting a PC subset from the plurality of PCs utilizing a random or pseudo-random selection algorithm, wherein in at least some of the PC sampling steps individual PCs are selected from the plurality of PCs with a probability that depends on current values of weight parameters assigned to each of the PCs;
- ii) computing an image relation model estimate for said PC subset;
- iii) computing, for each of the plurality of PCs, a PC error value for the image relation model estimate generated in said PC sampling step, and generating a score value for said image relation model estimate based on said PC error values for the plurality of PCs;
- and,
- iv) computing PC confidence values based on the PC error values of corresponding PCs;
- b) obtaining, with the processor, the output image relation model based on one of the image relation model estimates that has a best score value;
- and,
- c) saving the output image relation model in a computer-readable memory for generating new images;
- wherein at least some of the PC sampling steps further comprise updating values of the weight parameters based on cumulative PC confidence values computed for a sequence of PC sampling steps.
In a system comprising a processor and a memory for storing a plurality of point correspondences (PCs) between two images of a same scene or object obtained from one or more cameras, a method for determining an output image relation model relating the two images, comprising:
- a) using the processor to perform a series of PC sampling steps, each PC sampling step comprising:
- i) selecting a PC subset from the plurality of PCs utilizing a weighted random or pseudo-random selection algorithm wherein individual PCs are selected in dependence upon current values of weight parameters assigned thereto,
- ii) computing an image relation model estimate based on the PC subset;
- iii) computing, for each of the plurality of PCs, a PC error value for the image relation model estimate generated in said PC sampling step;
- and,
- iv) generating a score value for said image relation model estimate based on said PC error values for the plurality of PCs;
- b) obtaining, with the processor, the output image relation model using information related to one of the image relation model estimate with a best score;
- and,
- c) saving the output image relation model in a computer-readable memory for generating new images;
- wherein at least one of the PC sampling steps comprises:
- computing, by the processor, PC probability values based on the error values computed for corresponding PCs in a current PC sampling step and on a distribution of the error values computed for the plurality of PCs in the current PC sampling step, and
- updating by the processor the current values of the weight parameters using the PC probability values computed in said at least one of the PC sampling steps.
Dependencies are inferred from the claim text.