140 likes | 260 Views
Geometric Blur Descriptors for Point Correspondence. Nisarg Vyas Computational Photography (15862) Final Project, Carnegie Mellon University. Motivation. Point Correspondences are used in many vision applications Image Alignment
E N D
Geometric Blur Descriptors for Point Correspondence Nisarg Vyas Computational Photography (15862) Final Project, Carnegie Mellon University
Motivation • Point Correspondences are used in many vision applications • Image Alignment • 3D reconstruction of scene from multiple views • Object Recognition • Vehicle path Navigation • Structure from Motion
Point Correspondence Basic Approaches: SSD, NCC • ,Do not work well under affine transfoms
Blurred Descriptors • MOPS • Geometric Blur
Geometric Blur: Introduction • A “Spatially varying” Kernel which smoothes Instead of Kx(y) = Gσ(y), Kx(y) = Gα|x|(y)
Geometric Blur “Descriptor” • Take signed Gradient of input image in both directions, we will now be with 4 channels
Geometric Blur “Descriptor” Take a feature point, calculate Blur Descriptor for all 4 gradient channels, Subsampled in concentric circles
Status so far & Plans for final submission • Done implementing Geometric Blur Descriptor • Results are not as good as expected, sometimes simple SSD does even better !! • Have to try changing the thresholds which varies the sigma • Trying Other interesting descriptors (SIFT,C1), If time permits
References [1] Geometric Blur for Template Matching A.C. Berg and J. Malik, CVPR, 2001 [2] Shape Matching and Object Recognition using Low-distortion Correspondences, A.C. Berg, T.L. Berg and J. Malik, CVPR, 2005 [3] Comparing Visual Features for Morphing Based Recognition, J.J. Wu, MIT CSAIL Technical report, 2005 (TR-2005-035)