From: Xiaowei Li <ne...@bi...> - 2004-10-14 01:52:53
|
Hi,Everyone=A3=A1 =09I'm a new comer to VXL, and I wanna use it in my multiple view= geometry applications, since VXL seems having more expansibility= and easier implementation of many prevailing algorithms. I write= this mail to see whether you can give some ideas on my= experiments. =09One key issue in my experiment is to track a planar target in= image sequences. The plane would be identified at the first= frame, by drawing on that image a region inside which is rightly= the main portion of the plane. After this, the tracking is= absolutely automatic. Homography(H) between consecutive images= would be computed with RANSAC. According to the Multiple View= Geometry theory and the RANSAC method, this H's computation= would result in a robust estimation of H and a set of 2D point= correspondences as inliers, and of course these corresondences= are on the plane. "Extending" the region a bit more in the= previous image,making correspondences from the extended region= of the previous image to the consecutive image, and now using= the computed H to test these new correspondences, we could= verify which are the ones on the same plane, and furtherly a= rough boundary of that plane is obtained in the latter image.= Doing this in every frame, thus the plane is tracked robustly. =09That's the method. I think when we assume that no too large= occlusions enter that plane, this tracking would be sufficiently= robust. Do you think so? =09But I meet some problems in the implementation, because of my= unfamiliarness to the VXL environment. =09The first thing is to identify the plane by drawing a geometric= region in the 1st frame. The common geometric objects which= could represent the ROI in VXL are always simple rectangles= ,circles, or ellipses. How to represent more general objects in= VXL, such as regions needed in our experiment with free curves= as boundary, or polygons ? =09The second thing is to make correspondences. The first step to= use RANSAC is to get putative correspondences based on proximaty= and similiarity of their intensity neighbourhood. When we've= already got two sets of Harris points, one set for each image of= a image pair, how to match them ? Does there exist an= off-the-shelf class to achieve this in VXL ? =09Next thing is the robust estimation. I found that in the Robust= Estimation Library, the minimization of some objective functions= is also based on RANSAC sampling, not based on traditional= derivative analysis, Isn't it? Can this offer the same accuracy= ? =09The final thing is about the XCV. I think this is very useful= for me. I don't know whether anyone is still working on its= updation. If you do this work, can you mail me a newly version= of XCV, with source code,please ? =09Thank you very very much for your help ! =09=09 Best wishes, Shawrie Lee(Xiaowei Li) AR Lab, BIT, P.R. China ne...@bi... TEL:+86 10 6891 2565 2004-10-13 |