Generating complete depth maps in passive vision systems

The authors describe a hybrid system which uses multiple depth resources, rule-based depth interpolation and predictive rendering to generate accurate and dense depth maps. Initial depth estimates are obtained from multiple views (shape-from-generalised-stereo) and by shape-from-focus. The resultant sparse depth maps are then interpolated symbolically. The interpolated depth map for each region is corrected by inverse rendering. It is shown that nontrivial integration of depth resources reduces many of the problems normally associated with considering each depth resource individually