There are many different features that can be used to detect the objects, such as texture, shape, and color. Sicne they can be used altogether to get superior performance, we cannot say any of them is the best one. In this paper, the authors introduce a system which uses boundary fragment as the feature to enhance the capability of shape information. The reason that they don't use the entire object boundary as the feature is because of the limitation in edge detection. It's really hard to obtain a complete ojbect boundary from a noisy image (or because of the occlusion).
Like most of the object detectors, the BFM (Boundary-Fragment-Model) system needs a traning phase beforehand. The BFM system first detect edges by Canny detector from the training images. After finding the edges, several fragments are chosen accoding to the scores tested on the validation images. Since each fragment has the information of the object centroid, the system can estimate the object position from the matched fragments in the detection phase. The fragment candidates are used alone or combined together to form weak classifiers. From the weak classifiers, A strong classifier is obtained by the classic AdaBoost algorithm.
In the detection phase, the strong classifer is used to decide the existence of an object. If the object exists, Mean Shift is applied on the Hough voting space to get the centroid of the object. And then, the silhouette of this object is obtained by backprojection of each weak classifer.
In their experiment, they prove that their performance is competitive to other state-of-the-art detectors'. They also show their detector's ability to distingush between different objects with similar contours.
At the end of the paper, they discuss about the invariance to scale, rotation, and viewpoint. However, I think this part is explained ambiguously, and thus I have no more comments about this section.
Reference:
A. Opelt and A. Pinz and A. Zisserman, "A Boundary-Fragment-Model for Object Detection ." European Conference on Computer Vision, p. II: 575-588, 2006.
Like most of the object detectors, the BFM (Boundary-Fragment-Model) system needs a traning phase beforehand. The BFM system first detect edges by Canny detector from the training images. After finding the edges, several fragments are chosen accoding to the scores tested on the validation images. Since each fragment has the information of the object centroid, the system can estimate the object position from the matched fragments in the detection phase. The fragment candidates are used alone or combined together to form weak classifiers. From the weak classifiers, A strong classifier is obtained by the classic AdaBoost algorithm.
In the detection phase, the strong classifer is used to decide the existence of an object. If the object exists, Mean Shift is applied on the Hough voting space to get the centroid of the object. And then, the silhouette of this object is obtained by backprojection of each weak classifer.
In their experiment, they prove that their performance is competitive to other state-of-the-art detectors'. They also show their detector's ability to distingush between different objects with similar contours.
At the end of the paper, they discuss about the invariance to scale, rotation, and viewpoint. However, I think this part is explained ambiguously, and thus I have no more comments about this section.
Reference:
A. Opelt and A. Pinz and A. Zisserman, "A Boundary-Fragment-Model for Object Detection ." European Conference on Computer Vision, p. II: 575-588, 2006.
沒有留言:
張貼留言