On accurate dense stereo-matching using a local adaptive multi-cost approach

Defining pixel correspondences among images is a fundamental process in fully automating image-based 3D reconstruction. In this contribution, we show that an adaptive local stereo-method of high computational efficiency may provide accurate 3D reconstructions under various scenarios, or even outperform global optimizations. We demonstrate that census matching cost on image gradients is more robust, and we exponentially combine it with the absolute difference in colour and in principal image derivatives. An aggregated cost volume is computed by linearly expanded cross skeleton support regions. A novel consideration is the smoothing of the cost volume via a modified 3D Gaussian kernel, which is geometrically constrained; this offers 3D support to cost computation in order to relax the inherent assumption of “fronto-parallelism” in local methods. The above steps are integrated into a hierarchical scheme, which exploits adaptive windows. Hence, failures around surface discontinuities, typical in hierarchical matching, are addressed. Extensive results are presented for datasets from popular benchmarks as well as for aerial and high-resolution close-range images.