Method and system for estimating jointed-figure configurations

A method for determining the configuration of a limb on a figure, where the position of the limb can result from the combined movements of a plurality of jointed limbs, comprising the steps of:

• computing a first learned model which is based upon a first reference point that is determined with respect to a first limb of a figure, and using said first model to estimate the position of a second reference point on a second limb that is connected to said first limb;
• computing a second model which is based upon the second reference point, and using said second model to estimate the position of a third reference point relative to said second reference point;
• and
• combining the estimates for the second and third reference points to estimate the position of said third reference point relative to said first reference point.

In a system which estimates parameters of an image by means of a model which is based upon a data matrix comprising an input data matrix of image values and an output data matrix containing said parameters, a method comprising the steps of:

• (i) transforming the output data matrix by compressing variations in all rows of the output data matrix into a lesser number of rows, and applying the same transformation to the rows of the input data matrix;
• (ii) determining the relationship between (a) the coupling from each input dimension to the output variations in said lesser number of rows, and (b) the total variations in that input dimension;
• (iii) selecting one of said input dimensions on the basis of said determination;
• (iv) removing the selected input dimension and output variations which are coupled to the selected input dimension from the data matrix to produce a reduced data matrix;
• (v) iteratively repeating steps i-iv with respect to the reduced data matrix, to select additional input dimensions and coupled output variations;
• (vi) computing a model to estimate said parameters on the basis of said selected input dimensions and coupled output dimensions;
• and
• (vii) processing image values for an image whose parameters are unknown in accordance with said model to estimate the parameters for said image.

A method for estimating parameters which pertain to a visible image, comprising the steps of:

• developing an explicit hidden model which describes said parameters and variations of said parameters;
• using said hidden model to generate visual representations of images;
• training an eigen-points model on the basis of both said hidden model and said visual representations;
• generating a visual representation of an image whose parameters are unknown;
• and
• applying said visual representation to said eigen-points model to estimate said parameters for said image.

A method for estimating the configuration of a figure in an image, comprising the steps of:

• generating a plurality of poses for a figure, each having an associated configuration that is defined by a plurality of configuration parameters;
• for each pose, generating a signed distance image having values which identify the distance of pixels in the image from a border of the figure;
• computing a learned model for figure configurations, based upon said configuration parameters and said distance image values associated with said plurality of poses;
• generating a signed distance image for a new image of a figure whose configuration is unknown;
• applying the distance image values for said new image to said model;
• and
• generating a set of configuration parameters for the new image from said model.

A method for estimating the configuration of a figure in an image, comprising a learning phase and an estimation phase, wherein said learning phase includes the steps of:

• generating a plurality of poses for a figure, each having an associated configuration that is defined by a plurality of configuration parameters;
• for each pose, generating a representation image that is defined by a plurality of representation parameters;
• and
• computing a learned model for figure configurations, based upon said configuration parameters and said representation parameters associated with said plurality of poses;
• and wherein said estimation phase is carried out after the completion of said learning phase, and includes the steps of:
• generating a representation image for a new image of a figure whose configuration is unknown;
• applying the representation parameters for said new image to said model;
• and
• generating a set of configuration parameters for the new image from said model.

A method for determining the configuration of a limb on a figure, where the position of the limb can result from the combined movements of a plurality of jointed limbs, comprising the steps of:

• computing a first learned model which is based upon a first reference point that is determined with respect to a first limb of a figure, and using said first model to estimate a first region on a second limb that is connected to said first limb;
• determining a second reference point within said first region;
• computing a second model which is based upon the second reference point, and using said second model to estimate the position of a second region on the figure, relative to said second reference point;
• determining a third reference point within said second region;
• and
• combining the estimates for the first and second regions to estimate the position of said third reference point relative to said first reference point.