130 likes | 308 Views
Line Matching. Jonghee Park GIST CV-Lab. Introduction. Lines Fundamental feature in many computer vision fields 3D reconstruction, SLAM, motion estimation Useful features in low-textured and man-made structure Stereo Technology Assume that depth discontinuity mainly occurs nearby edges
E N D
LineMatching Jonghee Park GIST CV-Lab.
Introduction • Lines • Fundamental feature in many computer vision fields • 3D reconstruction, SLAM, motion estimation • Useful features in low-textured and man-made structure • Stereo Technology • Assume that depth discontinuity mainly occurs nearby edges • Applied vehicle, robot and aerial systems because of its low cost and depth information • To apply lines into practical stereo systems, computational complexity is critical issue for real-time performance.
Previous works • Line matching with pre-processing • Usually, point matching is conducted to find geometric relations between features and scenes • Scale and rotation [CVPR12], fundamental matrix and tri-pocal tensor [CVPR97] • Point matching takes long time because of feature extraction in scale space and construction of HOG • Line matching without pre-processing • Grouping based matching • Make groups or clusters lines for distinctive similarity measure according to the topological relation • LS[ICCV09] takes long time because of intensive 2D search for making multiple clusters • Individual matching • Match lines individually without topological relation • MSLD[PR12] makes multiple HOG for a line
MSLD: A robust descriptor for line matching Zhiheng Wang, Fuchao Wu, and Zhanyi Hu National Lab. Of Pattern Recognition PR 2009 Jonghee Park GIST CV-Lab.
Pixel Support Region Orientation
Sub-region Representation • Rotation relation of gradients between two images • Approximation • Weighting with gaussian kernel like SIFT • Interpolation along orientation direction for boundary effect Gradient Distance from line
Sub-region Representation • Each sub-region has following 4 dimension feature vector • Gradient description matrix • To cover line length variantion
Line Matching with Binary Costs Jonghee Park GIST CV-Lab.