Bibliography:
D. Sharon and M. van de Panne; EUROGRAPHICS Workshop on Sketch-Based Interfaces and Modeling (2006)
Summary:
This paper discusses an application of constellation models to develop probabilistic models for object sketches, based on multiple example drawings. These models are applied to estimate 'most-likely' labels for a sketch. The constellation model described in this paper is designed to capture the structure of a particular class of object and is based on local features and pairwise features, such as distances to other parts.
The probabilistic model is first learned from a set of labelled sketches. The recognition algorithm then determines a maximum-likelihood labelling for an unlabelled sketch by using a branch and bound algorithm. The particular method described in the paper allow considerable variability in the way sketches are drawn. However, the algorithm does make 2 assumptions in the way the sketches are drawn. 1) Similar parts are drawn with similar strokes. 2) Mandatory parts in an object are drawn exactly once.
The constellation model consists of 2 main feature vectors 1) The individual object part features 2) Pairwise features. In order to make the model efficient, pairwise features are calculated only form mandatory individual features. From the training examples, an object model is learned as a diagonal covariance matrix. The quality of a particular matching is measured using a cost function. A maximum likelihood search is performed to find the most plausible match. The search over all possible label assignments is carried out by a branch and bound search tree. Branches of a search are bound using multipart thresholding. Upon failure to find a label assignment, the process is repeated with a weaker threshold until a match is found.
The method was tested on 5 classes of objects, with 20-60 training examples each. The recognition time was found to be under 2.5 seconds, with most of the time spent on initialization. The multipass thresholding significantly reduced the computation time.
Discussion:
Constellation models seem like a nice way to identify complex shapes, that are domain specific. I like the fact that these do not depend on smaller basic shapes, that are usually required in a lot of gesture based methods. I am not sure how well this method would work when shapes are very similar. It'll be interesting to see how the threholds play out.
D. Sharon and M. van de Panne; EUROGRAPHICS Workshop on Sketch-Based Interfaces and Modeling (2006)
Summary:
This paper discusses an application of constellation models to develop probabilistic models for object sketches, based on multiple example drawings. These models are applied to estimate 'most-likely' labels for a sketch. The constellation model described in this paper is designed to capture the structure of a particular class of object and is based on local features and pairwise features, such as distances to other parts.
The probabilistic model is first learned from a set of labelled sketches. The recognition algorithm then determines a maximum-likelihood labelling for an unlabelled sketch by using a branch and bound algorithm. The particular method described in the paper allow considerable variability in the way sketches are drawn. However, the algorithm does make 2 assumptions in the way the sketches are drawn. 1) Similar parts are drawn with similar strokes. 2) Mandatory parts in an object are drawn exactly once.
The constellation model consists of 2 main feature vectors 1) The individual object part features 2) Pairwise features. In order to make the model efficient, pairwise features are calculated only form mandatory individual features. From the training examples, an object model is learned as a diagonal covariance matrix. The quality of a particular matching is measured using a cost function. A maximum likelihood search is performed to find the most plausible match. The search over all possible label assignments is carried out by a branch and bound search tree. Branches of a search are bound using multipart thresholding. Upon failure to find a label assignment, the process is repeated with a weaker threshold until a match is found.
The method was tested on 5 classes of objects, with 20-60 training examples each. The recognition time was found to be under 2.5 seconds, with most of the time spent on initialization. The multipass thresholding significantly reduced the computation time.
Discussion:
Constellation models seem like a nice way to identify complex shapes, that are domain specific. I like the fact that these do not depend on smaller basic shapes, that are usually required in a lot of gesture based methods. I am not sure how well this method would work when shapes are very similar. It'll be interesting to see how the threholds play out.
No comments:
Post a Comment