Authors: D. Sharon and M. van de Panne
Comments:
Summary:
This paper uses constellation model to recognize shapes. An object is defined with 2 features, individual and pair wise. Individual features give the shape and global location of the object. The pairwise features give the relative position of parts. Individual feature is a vector of X,Y( center of the line aligned to the bounding box), d(normalized length of the diagonal of the bounding box) and Cosine of the angle between the diagonal the X axis. Pairwise feature vector include difference in X, difference in Y, the minimum distance between the endpoints of stroke a and any point on stroke b, and the minimum distance between the endpoints of stroke b and any point on stroke a.
The sketch recognition process has two phases, the first which searches the space of possible mandatory label assignments, and the second which searches for optional labels for the remaining unlabelled strokes. In this way the mandatory labels provide contextual location information necessary for assigning appropriate labels to the potentially large number of optional parts.Mean and variance for the Individual and pair wise features are calculated separately and likelihood for labeling(L) is calculated by mutiplying the likelihood of individual features being recognized as L and likelihood of pair-wise features being labelled L.
A maximum likelihood search procedure is used to label the strokes. There are exponential number of possibilities of such selection. There are some strokes which might not be labelled. In order to reduce the computation cost because of the above mentioned factors, the search is done over 2 phases - labeling strokes which are mandatory and then linear search through the optional labels for the recognition of the remaining unlabeled strokes.The search is carried out using branch and bound tree. Multi pass thresholding is used to increase the speed of searching for large number of strokes.
Recognition can go wrong in several ways - inability to find suitable mandatory strokes because of the hard constraints, mislabeling of a mandatory stroke and mislabeling of optional strokes.
Errors are rare and imply a lack of training data, can occur if unusual strokes occur that affect
the overall bounding box and fewer mandatory strokes
Discussion:
The highlight of the paper is the constellation model. Using pairwise features to understand the spacial orientation is interesting. The number of training samples that this algorithm needs is quite large(20-60). The algorithm depends more on the bounding box and on the madatory strokes for the recognition. This is contrary to the flexibility it claims to give the user.
A maximum likelihood search procedure is used to label the strokes. There are exponential number of possibilities of such selection. There are some strokes which might not be labelled. In order to reduce the computation cost because of the above mentioned factors, the search is done over 2 phases - labeling strokes which are mandatory and then linear search through the optional labels for the recognition of the remaining unlabeled strokes.The search is carried out using branch and bound tree. Multi pass thresholding is used to increase the speed of searching for large number of strokes.
Recognition can go wrong in several ways - inability to find suitable mandatory strokes because of the hard constraints, mislabeling of a mandatory stroke and mislabeling of optional strokes.
Errors are rare and imply a lack of training data, can occur if unusual strokes occur that affect
the overall bounding box and fewer mandatory strokes
Discussion:
The highlight of the paper is the constellation model. Using pairwise features to understand the spacial orientation is interesting. The number of training samples that this algorithm needs is quite large(20-60). The algorithm depends more on the bounding box and on the madatory strokes for the recognition. This is contrary to the flexibility it claims to give the user.
No comments:
Post a Comment