Two lines of parallel translation of PMVS algorithm

Abstract

Sparse 3D reconstruction by using the incremental motion recovery structure system. First, SIFT feature points in the English text sequence were extracted, and mismatches were removed by reverse screening method and RANSAC algorithm. According to the deficiency of PMVS algorithm in the reconstruction process, the corresponding improvement method is proposed. The PMVS algorithm was first used to obtain a rough quasi-English two-line parallel translation system, The projection matching points of the point cloud are obtained through the projection matrix, Then, the method based on the proximity point distance constraint, ZNCC stereo matching constraint and the pole line constraint is used for the regional diffusion of the matching points; Then use the template matching algorithm to obtain the corresponding matching block of the point cloud hole on two lines of parallel translated English text, The ZNCC stereo matching algorithm with the adaptive window size was used to obtain the matching points within the matching block, Finally, the spatial points corresponding to the matching points are obtained by sub-pixel interpolation and triangulation, Finally, a two-line parallel translation system is reconstructed. Classified the Chinese and English sentences into simple short sentences and complex long sentences. For simple short sentences, the rules-based and statistical methods are used to align the more complex long sentences, and then align the short sentences. In the phrase recognition stage, the Chinese-English bilingual "marker words" set is used to cut the Chinese-English sentences to obtain the "marker words" phrase. Then, the basic noun phrases were identified using a bilingual corpus-based approach. In the Temple dataset and Dino dataset, this paper proposes that the improved PMVS algorithm has 11.11 % and 10.64 % improvement in time efficiency compared to the original PMVS algorithm. The time used by the two algorithms in the first stage is given. According to the data in the table, for the data set Temple, the original algorithm takes 49 s, while the improved PMVS algorithm takes 85 s, which takes more time than the original algorithm.