Conventional stereo correspondence algorithms which search for a ?best? match are defeated by the many sources of noise possible in a pair of stereo images. We propose a new reconstruction paradigm, Concurrent Stereo Matching (CSM), that starts with a noise model and marks regions which could not be considered matches - given the noise model. The work presented here uses spatially varying noise values obtained empirically from segmented images. These noise levels determine admissable matches and define candidate surfaces, which are then processed using local constraints only to a final set of reconstructed surfaces. For a complex scene with many small surfaces, CSM ranks highly amongst existing benchmarked algorithms. The current CSM implementation does not handle large sloping surfaces well but work is underway to rectify this.