Parallel refinement of slanted 3D reconstruction using dense stereo induced from symmetry

Traditional dense stereo estimation algorithms measure photo-similarity to calculate the disparity between image pairs. SymStereo is a new framework of matching cost functions that measure symmetry to evaluate the possibility of two pixels being a match. This article proposes a fully functional real-time parallel 3D reconstruction pipeline that uses dense stereo-based photo-symmetry. The logN variant of SymStereo achieves superior results for images with slanted surfaces, when compared with other algorithms (Antunes and Barreto in Int J Comput Vis 1-22, 2014). This is of particular interest for areas of computer vision such as the processing of datasets for urban scene reconstruction and also for tracking in robotics or intelligent autonomous vehicles. The output results obtained are analyzed by tuning distinct matching cost, aggregation and refinement parameters, targeting the most suitable combinations for slant dominated images. Also, the parallel approach for the aforementioned pipeline consists of a hybrid dual GPU system capable of calculating from 2 up to 132 volumes per second for high- and low-resolution images, respectively.