Sparse Representation with Geometric Configuration Constraint for Line Segment Matching

Abstract

We present a novel line segment matching method based on sparse representation with geometric configuration constraint. The significant idea is that we transfer the line matching issue into sparsity based line recognition. At first, line segments are detected by LSD detector and clustered according to spatial proximity to form completed lines. SIFT is used to represent points in the line segments and all point features are put together to form a distinctive descriptor. Line feature is then represented by a max pooling function. Features of all line segments are trained into a dictionary using sparse coding. Lines with the same similarity may fall together in the high dimensional feature space. Finally, lines in one view are matched to their counterparts in other views by seeking pulses from the coefficient vector. Under our framework, line segment is trained once and can be matched over all other views. When compared to matching approaches based on local invariant features, our method shows encouraging results with high efficiency. Experiment results have validated the effectiveness for planar structured scenes under various transformations.