System and method for automatic landmark labeling with minimal supervision
AssigneeNBCUniversal Media, LLC
InventorsYan Tong, Xiaoming Liu, Jilin Tu, Peter Henry Tu, Frederick Wilson Wheeler
Life of this patent
Estimating the term…
Abstract
A system and method for estimating a set of landmarks for a large image ensemble employs only a small number of manually labeled images from the ensemble and avoids labor-intensive and error-prone object detection, tracking and alignment learning task limitations associated with manual image labeling techniques. A semi-supervised least squares congealing approach is employed to minimize an objective function defined on both labeled and unlabeled images. A shape model is learned on-line to constrain the landmark configuration. A partitioning strategy allows coarse-to-fine landmark estimation.
Claims20
A method comprising:
- via a computer processor:
- identifying a whole warped image as a first level patch;
- obtaining initial landmark locations from the first level patch;
- iteratively partitioning the whole warped image region into smaller child patches;
- and
- refining landmark estimations from the initial landmark locations to establish accurate landmark labeling based on resultant patch appearance.
A method comprising:
- via a computer processor:
- performing a semi-supervised least-squares-based alignment of an image ensemble using an inverse warping technique;
- determining estimated warping parameters of unlabeled images in the image ensemble based on known warping parameters of labeled images of the image ensemble using results of the semi-supervised least-squares-based alignment;
- reducing outliers of the estimated warping parameters by partitioning a mean shape space of the semi-supervised least-squares-based alignment, wherein partitioning the means shape space includes iteratively partitioning an initially identified patch corresponding to an image of the image ensemble into child patches;
- and
- refining landmark estimations based on resultant patch appearance.
A tangible, non-transitory, computer readable medium comprising machine-readable instructions to:
- perform a semi-supervised least-squares-based alignment of an image ensemble using an inverse warping technique;
- determine estimated warping parameters of unlabeled images in the image ensemble based on known warping parameters of labeled images of the image ensemble using results of the semi-supervised least-squares-based alignment;
- reduce outliers of the estimated warping parameters by partitioning a mean shape space of the semi-supervised least-squares-based alignment, wherein partitioning the means shape space includes iteratively partitioning an initially identified patch corresponding to an image of the image ensemble into child patches;
- and
- refine landmark estimations based on resultant patch appearance.
Dependencies are inferred from the claim text.