Multi-view Stereo Reconstruction via Homogeneous Spatial Expansion

A multi-view stereo reconstruction algorithm based on homogeneous spatial expansion was proposed to overcome the difficulty of reconstructing3D models with both high accuracy and high completeness for scenes in complex situations.It expanded the frame work of the feature expansion algorithm with a new conditional initial value rectification step.Meanwhile the existing steps were either replaced or improved.Firstly,sparse seed points were extracted by SFM method.Then they were expanded simultaneously to their nearby3D space directly in a fixed global world coordinates system to gain initial positions and normals for growing points.The final positions and normals of the growing points were determined via optimization,during which the best reference and auxiliary views were selected and replaced according to the current status of the growing points.Next,to avoid converging into local minima and improve the reconstruction accuracy,the positions and normals of grown points were conditionally rectified taking advantage of their neighbor points.Once rectified,re-optimization was carried out with new initial values.Finally,to suppress the spreading of errors the adaptive consistent filters for smoothness,depth and direction were designed to remove outliers and keep accurate points as far as possible.The experiments showed that both the reconstruction accuracy and completeness of the proposed algorithm are obviously higher than that of the popular PMVS algorithm.With much improved completeness,its accuracy is still on a par with the high accuracy of DAISY algorithm.