Parameterized model of 2D articulated human shape
A novel “contour person” (CP) model of the human body is proposed that has the expressive power of a detailed 3D model and the computational benefits of a simple 2D part-based model. The CP model is learned from a 3D model of the human body that captures natural shape and pose variations; the projected contours of this model, along with their segmentation into parts forms the training set. The CP model factors deformations of the body into three components: shape variation, viewpoint change and pose variation. The CP model can be “dressed” with a low-dimensional clothing model, referred to as “dressed contour person” (DCP) model. The clothing is represented as a deformation from the underlying CP representation. This deformation is learned from training examples using principal component analysis to produce so-called eigen-clothing. The coefficients of the eigen-clothing can be used to recognize different categories of clothing on dressed people. The parameters of the estimated 2D body can be used to discriminatively predict 3D body shape using a learned mapping approach. The prediction framework can be used to estimate/predict the 3D shape of a person from a cluttered video sequence and/or from several snapshots taken with a digital camera or a cell phone.
1. A method, comprising:
generating a three-dimensional model of an unclothed human body, the three-dimensional model capturing one or more shapes or poses of the unclothed human body;
determining two-dimensional contours associated with the three-dimensional model;
segmenting the two-dimensional contours into one or more segments;
based on the one or more segments, generating a two-dimensional model of the unclothed human body, the two-dimensional model factoring deformations of the unclothed human body into a shape variation component, a viewpoint change, and a pose variation; and
generating a second two-dimensional model is of a clothed human body based on the two-dimensional model of the unclothed human body, wherein generating the second two-dimensional model comprises:
computing deformations by aligning a contour of the clothed human body with a second contour of the unclothed human body; and
learning an eigen-clothing model using principal component analysis applied to the deformations, wherein the eigen-clothing model classifies different types of clothing.
2. The method of claim 1 , wherein the two-dimensional model visualizes frontal views and non-frontal views of the unclothed human body.
3. The method of claim 1 , wherein the pose variation comprises at least one of a body parts rotation and foreshortening.
4. The method of claim 1 , wherein the deformations of the unclothed human body comprise non-rigid deformations resulting from articulation.
5. The method of claim 1 , wherein the two-dimensional model is factored into a linear approximation to distortions caused by local camera view changes.
6. The method of claim 1 , wherein the two-dimensional model is factored into an articulation of body parts represented by a rotation and length scaling.
7. The method of claim 1 , wherein the two-dimensional model is factored into a linear model characterizing shape changes across a population.
8. The method of claim 1 , wherein generating the second two-dimensional model comprises defining a set of linear coefficients that produce from the unclothed human body a particular deformation associated with a clothing type.
9. A non-transitory computer-readable storage device having stored therein instructions which, when executed by a processor, cause the processor to perform operations comprising:
generating a three-dimensional model of an unclothed human body, the three-dimensional model capturing one or more shapes or poses of the unclothed human body;
determining two-dimensional contours associated with the three-dimensional model;
segmenting the two-dimensional contours into one or more segments;
based on the one or more segments, generating a two-dimensional model of the unclothed human body, the two-dimensional model factoring deformations of the unclothed human body into a shape variation component, a viewpoint change, and a pose variation; and
generating a second two-dimensional model is of a clothed human body based on the two-dimensional model of the unclothed human body, wherein generating the second two-dimensional model comprises:
computing deformations by aligning a contour of the clothed human body with a second contour of the unclothed human body; and
learning an eigen-clothing model using principal component analysis applied to the deformations, wherein the eigen-clothing model classifies different types of clothing.
10. The non-transitory computer-readable storage device of claim 9 , wherein the two-dimensional model visualizes frontal views and non-frontal views of the unclothed human body, and wherein the pose variation comprises at least one of a body parts rotation and foreshortening.
11. The non-transitory computer-readable storage device of claim 9 , wherein the deformations of the unclothed human body comprise non-rigid deformations resulting from articulation, and wherein the two-dimensional model is factored into an articulation of body parts represented by a rotation and length scaling.
12. The non-transitory computer-readable storage device of claim 9 , wherein generating the three-dimensional model of the unclothed human body comprises randomly sampling a body shape from a plurality of body shapes, wherein the body shape is sampled in a random pose from a plurality of poses and viewed from a random camera from a plurality of cameras.
13. The non-transitory computer-readable storage device of claim 9 , wherein the two-dimensional model is factored into a linear model characterizing shape changes across a population.
14. The non-transitory computer-readable storage device of claim 9 , wherein generating the second two-dimensional model comprises defining a set of linear coefficients that produce from the unclothed human body a particular deformation associated with a clothing type.
15. A system comprising:
a processor; and
a non-transitory computer-readable storage medium having stored therein instructions which, when executed by the processor, cause the processor to perform operations comprising:
obtaining a 3D model of an unclothed human body that captures shape and pose variations;
obtaining a training set associated with the 3D model, the training set comprising projected contours and a segmentation of the contours into parts;
based on the training set, generating a 2D contour person (CP) model of the unclothed human body, wherein the 2D CP model comprises a shape variation component, a viewpoint change component, and a pose variation component; and
generating a second 2D model is of a clothed human body based on the 2D CP model of the unclothed human body, wherein generating the second 2D model comprises:
computing deformations by aligning a contour of the clothed human body with a second contour of the unclothed human body; and
learning an eigen-clothing model using principal component analysis applied to the deformations, wherein the eigen-clothing model classifies different types of clothing.
16. The system of claim 15 , wherein obtaining the 3D model of the unclothed human body comprises randomly sampling a body shape from a plurality of body shapes, wherein the body shape is sampled in a random pose from a plurality of poses and viewed from a random camera from a plurality of cameras.
17. The system of claim 16 , wherein the projected contours are generated by projecting a 3D body contour onto a camera plane to produce a training contour.