Feature Matching Based on Grid and Multi-Density for Ancient Architectural Images

The 3 D reconstruction of Chinese ancient architecture from images is a difficult problem in image based modeling, where the image matching is a key factor. To improve the feature matching performance cross images, a new method, FM_GMC, is proposed in this work, which is based on grid partition and multi-density clustering of key-points in grids. Firstly, SIFT is adopted to extract images key-points. Then images are partitioned into grids and key-points densities in grids are computed, from which anchor cells, neighbor cells, and border cells are determined. Meanwhile, the connected regions of such labeled cells are defined as clusters according to the similarity of local regions. Finally, key-points matching within similar clusters is carried out by the nearest neighbor distance ratio(NNDR). The matching results validate our proposed method on 3 D Reconstruction Dataset and 141 typical architectural images about Jinci.