Title : Shape Matching and Object Recognition Using Shape Contexts
Author : S. Belongie, J. Malik, and J. Puzicha
Summarization:
This paper proposed a shape descriptor : shape context, to describe the coarse distribution of the rest of the shape to a point on the shape. And solve the correspondence problem by computing sum of matching errors between corresponding points to find the nearest neighbor.
In their definition, a shape is a set of edge pixels which are found by edge detector. For computing the distance between two shape, they used the shape context which compute the coarse histogram of n-1 points in the relative coordinates position by the query point, and used this histogram to compute the difference with the other shape to find the best match points-pair with the other shape.
2009年3月24日 星期二
2009年3月10日 星期二
paper critique & summarization : Nonlinear Dimensionality Reduction by Locally Linear Embeddign
Title : Nonlinear Dimensionality Reduction by Locally Linear Embedding.
Author : S. T. Roweis and L. K. Saul
summarization:
This paper developed a unsupervised learning algorithm that could be used for dimension reduction.
The based ideas are:
1. Find the neighbors for each points in original coordinate system(maybe by KNN).
2. find some point in the neighbors could reconstruct the original point by linear combination.
3. map all the information above onto the low dimensions coordinate system and minimize the reconstruction error.
critique:
this approach is easy to figure out, but I like the figure 1 he draw. For PCA or MDS, they also directly project the data onto the low dimension system, so some far data would be projected into near position as the figure show. And this method use the neighbor concept to avoid this situation, it easy but make sense, I like this figure.
Author : S. T. Roweis and L. K. Saul
summarization:
This paper developed a unsupervised learning algorithm that could be used for dimension reduction.
The based ideas are:
1. Find the neighbors for each points in original coordinate system(maybe by KNN).
2. find some point in the neighbors could reconstruct the original point by linear combination.
3. map all the information above onto the low dimensions coordinate system and minimize the reconstruction error.
critique:
this approach is easy to figure out, but I like the figure 1 he draw. For PCA or MDS, they also directly project the data onto the low dimension system, so some far data would be projected into near position as the figure show. And this method use the neighbor concept to avoid this situation, it easy but make sense, I like this figure.
paper critique & summarization : Eigenfaces for Recognition
Title : Eigenfaces for Recognition
Author : Matthew Turk and Alex Pentland
summarization:
This paper developed a method to recognize human face based on projecting new faces onto the feature space, rather then depend on the 3-D information or geometry which before researches used, and then called these feature space "eigenface".
The idea is com from a technique developed by Sirovich and Kirby(1987&1990), they represented pictures of faces by using principal component analysis. They calculated a coordinate system for image compression and each coordinate is an image, they called them "eigenpicture".
And this approach has some base intuitions:
1. every human faces have some similarities, like the rough positions of eyes, nose, and mouth.
2. we can compute a average face, which would be the average of all training faces.
3. compute the differ of average face and all training faces, and keeping the M faces which correspond to the highest eigenvaluse, defined them as "face space".
4. then every face could be combined by using this eigenfaces with different weight.
critique:
This paper is similar to the other paper we read this week, they both use some component to combine the original data. And the number of component used also decides how precise the combined data with the original data. And by selecting few principal component, we could get the not bad performance and use little dimension to store the information, to get the effect of dimension reduction we want.
Author : Matthew Turk and Alex Pentland
summarization:
This paper developed a method to recognize human face based on projecting new faces onto the feature space, rather then depend on the 3-D information or geometry which before researches used, and then called these feature space "eigenface".
The idea is com from a technique developed by Sirovich and Kirby(1987&1990), they represented pictures of faces by using principal component analysis. They calculated a coordinate system for image compression and each coordinate is an image, they called them "eigenpicture".
And this approach has some base intuitions:
1. every human faces have some similarities, like the rough positions of eyes, nose, and mouth.
2. we can compute a average face, which would be the average of all training faces.
3. compute the differ of average face and all training faces, and keeping the M faces which correspond to the highest eigenvaluse, defined them as "face space".
4. then every face could be combined by using this eigenfaces with different weight.
critique:
This paper is similar to the other paper we read this week, they both use some component to combine the original data. And the number of component used also decides how precise the combined data with the original data. And by selecting few principal component, we could get the not bad performance and use little dimension to store the information, to get the effect of dimension reduction we want.
2009年3月3日 星期二
paper critique & summarization : Distinctive Image Features from Scale-Invariant Keypoints
Title: Distinctive Image Features from Scale-Invariant Keypoints
Author: D. G. Lowe
This paper proposes a approach named "Scale Invariant Feature Transform(SIFT)", which transforms image data into scale-invariant coordinates relative to local features.
the approach has 4 major stages, which introduced as follow:
1. Scale-space extrema detection - uses the DoG function to find the potential interest points, by comparing the 26 neighbor points and be selected if the DoG is larger or smaller then all others.
2. Keypoint localization - a detailed model is fit to determine location and scale. To delete the candidate points which contrast is low or on the edge.
3. Orientation assignment - one or more orientations are assigned to each keypoint location based on local image gradient directions. And the image data are transformed by this orientation to make sure the invariant of rotation.
4. Keypoint descriptor - represents these points with the parameters of location, scale, and orientation which assigned to these points.
And the keypoint descriptors generated by the above stages are highly distinctive, which is invariant in rotation, scale, and viewpoint.
Author: D. G. Lowe
This paper proposes a approach named "Scale Invariant Feature Transform(SIFT)", which transforms image data into scale-invariant coordinates relative to local features.
the approach has 4 major stages, which introduced as follow:
1. Scale-space extrema detection - uses the DoG function to find the potential interest points, by comparing the 26 neighbor points and be selected if the DoG is larger or smaller then all others.
2. Keypoint localization - a detailed model is fit to determine location and scale. To delete the candidate points which contrast is low or on the edge.
3. Orientation assignment - one or more orientations are assigned to each keypoint location based on local image gradient directions. And the image data are transformed by this orientation to make sure the invariant of rotation.
4. Keypoint descriptor - represents these points with the parameters of location, scale, and orientation which assigned to these points.
And the keypoint descriptors generated by the above stages are highly distinctive, which is invariant in rotation, scale, and viewpoint.
paper critique & summarization : Scale & Affine Invariant Interest Point Detectors
Title : Scale & Affine Invariant Interest Point Detectors
Author: K. Mikolajczyk and C. Schmid
This paper proposes a method for detecting interest points invariant to scale and affine transformations. It combines the Harris detector with Laplacian to get the goal.
Harris detector is a interest point detector, but it is not invariant to scale changes. To solve this problem, they generate multi-scale for each point, and use Laplacian-of-Faussians to find the highest percentage of correct characteristic scales. Then can generate a scale invariant detector, and they call it "Harris-Laplace detector".
And for the affine invariant interest point detector, because the Harris-Laplace detector could not reflect the real transformation of a point when it has not only scale change but also affine transformation. They use the "Second Moment Matrix" to fix the problem. But in fact, the math formulas is hard for me.
And the "Harris-Affine Interest Point Detector" which they propose have the following parts:
1. spatial localization - determined by the local maximum of the Harris function.
2. integration scale - selected at the extremum over scale of the normalized Laplacian.
3. differentiation scale - selected at the maximum of normalized isotropy
4. shape adaptation matrix - estimated with the second moment matrix.
Author: K. Mikolajczyk and C. Schmid
This paper proposes a method for detecting interest points invariant to scale and affine transformations. It combines the Harris detector with Laplacian to get the goal.
Harris detector is a interest point detector, but it is not invariant to scale changes. To solve this problem, they generate multi-scale for each point, and use Laplacian-of-Faussians to find the highest percentage of correct characteristic scales. Then can generate a scale invariant detector, and they call it "Harris-Laplace detector".
And for the affine invariant interest point detector, because the Harris-Laplace detector could not reflect the real transformation of a point when it has not only scale change but also affine transformation. They use the "Second Moment Matrix" to fix the problem. But in fact, the math formulas is hard for me.
And the "Harris-Affine Interest Point Detector" which they propose have the following parts:
1. spatial localization - determined by the local maximum of the Harris function.
2. integration scale - selected at the extremum over scale of the normalized Laplacian.
3. differentiation scale - selected at the maximum of normalized isotropy
4. shape adaptation matrix - estimated with the second moment matrix.
訂閱：
文章 (Atom)