construct texture description 基础知识.ppt

Construct texture description 实 验 实 验 致谢 向被我打扰过的洪晓鹏师兄、李安南师兄、翟德明师姐和阚美娜师姐表示感谢 感谢各位在这听我“瞎扯” * Figure 2. Left: Original image and first 3 dimensions of the embedding. Second and third columns: similarity of sampled image patches (blue dots) with regard to a selected patch shown in red,brightness is proportional to similarity. Proximity in the embedding space is equivalent to image-space similarity. * Figure 2. Left: Original image and first 3 dimensions of the embedding. Second and third columns: similarity of sampled image patches (blue dots) with regard to a selected patch shown in red,brightness is proportional to similarity. Proximity in the embedding space is equivalent to image-space similarity. * Figure 5. Tracking of a lightly textured object on a heavily cluttered background. Left Column: Unconstrained matching between image pairs. The red rectangles on the reference image are hand drawn and those on the target image are transformed by the homography computed using the matching key points. Because of large amounts of background clutter and erroneous correspondences the homography is bad. Right Column: Our approach to matching. The correspondences are much better and background is mostly eliminated. This is evidenced by the fact that the transformed rectangles now match much more accurately the car’s location. Bottom Row: Keypoint clusters produced by our algorithm for the four target images. Keypoints assigned to the foreground are shown in green and the rest in red. * Figure 6. Tracking a cheetah against a complex background. Left column: Standard unconstrained matching between a video frame and the reference image that appears in the top-left corner of Fig.1. Middle column: Foreground/background clusters produced by our method for the target video frame. Right column: Constrained matching results. * * 陈 静 2009.09.04 JDL视觉建模与识别组 文献列表 标题：Combining powerful local and global statistics for texture description 作者：Y