Japanese / English

Detail of Publication

Text Language English
Authors Fairuz Safwan Mahad, Masakazu Iwamura, Yuzuko Utsumi, Koichi Kise
Title Denser Feature Correspondences for 3D Reconstruction
Journal IPSJ SIG-CVIM: Computer Vision and Image Media
Vol. 2016-CVIM-202
No. 3
Pages pp.1-5
Location Ritsumeikan University
Reviewed or not Not reviewed
Presentation type Oral and Poster
Month & Year May 2016
Abstract The 3D reconstruction of an object is heavily dependent on the amount of texture on the surface of the object and the correspondences acquired between images from different viewpoints. Objects with lack of texture affects the reconstructed 3D model degrading its quality and completeness. This occurs due to the SIFTs and Harris's incapability of extracting sufficient features from regions with less texture thus resulting in a lack of correspondences in such regions. Gaps occurs in the reconstructed 3D model due to the lack of correspondences. The paper aims to obtain more correspondences through the implementation of convolution in conjunction with a set of randomized kernels. The use of randomized kernels changes the structure of the images in different ways allowing SIFT to extract features in less textured regions which could not be realized conventionally. Experimental results showed that the proposed method is able to fill up more gaps improving the completeness of the 3D model of less textured objects.
URL http://id.nii.ac.jp/1001/00159114/
Back to list