Effective seed generation for 3D reconstruction

Given a pair of calibrated cameras, we describe an effective seed construction method for 3D reconstruction which starts with initial estimates of seed position, improves them and computes good estimates normals. We formulate the seed construction as an optimization problem with a criterial function based on the similarity of reprojection of images on a hypothetical planar patch. We show that the criterial function is unimodal in certain area and all values in this area are greater than the values outside of this area. The ability to estimate seeds depends on the surface texture. Some methods evaluate the variance of intensities of seed texture to decide about the possibility of normal detection. We show that there also exist nonhomogenous textures which are not discriminative. Our method is able to detect situations when the seed normal is not possible to detect e.g. the texture is not discriminative. The method found global optimum in 94% of our test data. We show in experiments, that our approach outperforms other most relevant approaches.