Methods were proposed to estimate a surface from a sparse cloud of points reconstructed from images. These methods are interesting in several contexts including large scale scenes, limited computational resources and initialization of dense stereo. However they are deficient in presence of thin structures such as posts, which are often present in both urban and natural scenes: these scene components can be partly or even completely removed. Here we reduce this problem by introducing a pre-processing, assuming that (1) some of the points form polygonal chains approximating curves and occluding contours of the scene and (2) the direction of the thin structures is roughly known (e.g. vertical). In the experiments, our pre-processing improves the results of two different surface reconstruction methods applied on videos taken by helmet-held 360 cameras.