Wednesday, May 18, 2005
Diffeomorphic matching and image statistics
Taken from http://www.enpc.fr/certis/Papers/05certis05.pdf
By: Jean-Yves Audibert, Guillaume Charpiat, Olivier Faugeras, Renaud Keriven
Overview:
The research finds it necessary to use statistics on a set of 2D images to identify a person. This approach does not require any manual intervention to identify landmarks or regions of interest. They work directly on the deformation fields, which establish the correspondences between the whole images (this is the fundamental elements of the problem). Here, they deployed non-supervised algorithms to build such correspondence fields between images. Matching methods that were employed in his research is based on Gerardo’s (2002).
Compute (a) 1st order statistics
(b) 2nd order statistics
(c) Classification for recognition purposes – SVM with Gaussian kernel.
Image matching
The tasks of integrating information from different types of sensors, comparing data acquired at different time and/or putting the similar structure of two different images into correspondence are also known as image registration and template matching, and they are based upon the ability to automatically map points between the respective domains of images (i.e. image matching).
Given two images, A and B. The images are defined in a rectangular subject W of the plane R2. A deformation field, f, is calculated so that the warp image A o f resembles B. The deformation field should be smooth and invertible, meaning that the diffeomorphism f from the rectangular subset W to the identity on the image boundary. Then a criterion is chosen. A criterion expresses the similarity between the 2 images A and B. Hence, this project employed image matching algorithm which consists of minimising with respect to the deformation field through a multi-scale gradient descent.
The first order statistics of a set of images were computed with a diffeomorphic matching approach (without landmark) and then use them in a classification task. They are still in the process of including second order statistics in the classification algorithm.
Experiment and results:
To measure the efficiency of the method, we cross-validate the errors by taking out one subject among the 13 subjects in the database and consequently using 60 faces labeled by their expression to deduce the expression of the five remaining faces. The cross-validation error is 24 upon 65 faces.
For comparison purposes, we trained a Support Vector Machine with Gaussian kernel using only the gray level intensity information. In this case we obtained a larger cross-validation error of 27 upon 65 faces which shows the interest of using the diffeomorphisms.
Critical discussion:
The results can be encouraging. However, from the experiments, there is no drastic improvement. Thorough analysis and further enhancement required to improve the proposed methods used.
By: Jean-Yves Audibert, Guillaume Charpiat, Olivier Faugeras, Renaud Keriven
Overview:
The research finds it necessary to use statistics on a set of 2D images to identify a person. This approach does not require any manual intervention to identify landmarks or regions of interest. They work directly on the deformation fields, which establish the correspondences between the whole images (this is the fundamental elements of the problem). Here, they deployed non-supervised algorithms to build such correspondence fields between images. Matching methods that were employed in his research is based on Gerardo’s (2002).
Compute (a) 1st order statistics
(b) 2nd order statistics
(c) Classification for recognition purposes – SVM with Gaussian kernel.
Image matching
The tasks of integrating information from different types of sensors, comparing data acquired at different time and/or putting the similar structure of two different images into correspondence are also known as image registration and template matching, and they are based upon the ability to automatically map points between the respective domains of images (i.e. image matching).
Given two images, A and B. The images are defined in a rectangular subject W of the plane R2. A deformation field, f, is calculated so that the warp image A o f resembles B. The deformation field should be smooth and invertible, meaning that the diffeomorphism f from the rectangular subset W to the identity on the image boundary. Then a criterion is chosen. A criterion expresses the similarity between the 2 images A and B. Hence, this project employed image matching algorithm which consists of minimising with respect to the deformation field through a multi-scale gradient descent.
The first order statistics of a set of images were computed with a diffeomorphic matching approach (without landmark) and then use them in a classification task. They are still in the process of including second order statistics in the classification algorithm.
Experiment and results:
To measure the efficiency of the method, we cross-validate the errors by taking out one subject among the 13 subjects in the database and consequently using 60 faces labeled by their expression to deduce the expression of the five remaining faces. The cross-validation error is 24 upon 65 faces.
For comparison purposes, we trained a Support Vector Machine with Gaussian kernel using only the gray level intensity information. In this case we obtained a larger cross-validation error of 27 upon 65 faces which shows the interest of using the diffeomorphisms.
Critical discussion:
The results can be encouraging. However, from the experiments, there is no drastic improvement. Thorough analysis and further enhancement required to improve the proposed methods used.
# posted by Jacey Lynn Minoi : 4:27 pm
