Merging depth maps - A study in how a varying number of depth maps affect the results

Abstract: A depth map is normally created using two horizontally displaced images, where each depth value in the map is the distance to the 3D point. In some image areas, depth values cannot be calculated due to low texture, occlusions etc. These areas are classified as uncertain. This thesis proposes a simple method using multiple depth maps to fill in uncertain areas and improve the depth estimates elsewhere. The accuracy and coverage of the results are evaluated as a function of the number of depth maps used. The method was evaluated on three data sets from the Middlebury data set. In general, the merged maps have better accuracy and coverage thantheindividualdepthmaps. Furthermore,theaccuracyandcoverage improves with an increasing number of individual maps. The exception is a narrow transitory coverage range where the accuracy worsens with increasing number of individual depth maps.