Improving PMVS Based On Quasi-dense Matching

With the development of computer technology,3D reconstruction becomes aresearch topic of the computer visions and has a very wide application prospects inthe fields of medical diagnosis, visual navigation, movie,3D games, etc. Multi-viewbased reconstruction is the mainly method to reconstruct3D models and manyexcellent algorithms has been reported these years, such as the Patch-basedMulti-View Stereo (PMVS) by Furukawa. This thesis provides a deep research on thePMVS and proposes an improved PMVS algorithm.PMVSuses the Harris and DoG operators to detect features, so the set of initialpatches generated by the feature matching step is sparse, leading to massive work ofthe expansion step and spending a lot of time. To solve this problem, this thesisintroducesthe method of quasi-dense matching, and proposes the PMVS algorithmbased on quasi-dense matching. The experiment demonstrates that this improvedalgorithm can reduce the total computation time as expected, but the non-uniformdistribution of initial patches produced by the quasi-dense matching approach resultsin declining of completeness of the reconstruction results.Two simple but effectivemethods are proposed to handle this problem. First, this thesis enforces homographconstraints in the feature matching step and adopts the adaptive non-maximalsuppression method to handle the sparse matches to enhance the reliability of thequasi-dense matches. Second,we use resampling approach to deal with thequasi-dense matched points to ensure the uniform distribution of the initial patches,thus the initial patches can expansion effectively which ensures the completeness offinal reconstruction result to be approximately equal with the original PMVS.Experiment results demonstrate that the improved algorithm proposed by thisthesis have8.85%time saved on Temple dataset and9.21%time saved on Dinodataset, that means the improved algorithm has improved the efficiency of the3dreconstruct. Also we submitted the reconstruct results to the Middlebury benchmark toevaluate the effect, the evaluation shows the accuracy of the improved PMVS is0.9mm and0.54mm on Temple and Dino dataset respectively, the completeness of theimproved PMVS is97.8%and98.4%. This results show that the improved algorithmhas equal effect with the original PMVS and the reconstruct results can be accepted.