Abstract:  The PMVS (patch-based multi-view stereo) algorithm is widely used in the multi-view stereo field because of its good performance. However, the surface details of the reconstructed model are lost and the exact location of the reconstruction points is not accurate. These problems are very serious in some cases, especially while the input images are few and the texture of the reconstructed model is less. In order to solve these problems, this paper studies the removal of mismatched candidate points and the ranking of seed point reliability: firstly, use USAC (Universal-RANSAC) to remove the candidate matching error; then, put forward the double constraint strategy, meanwhile select the candidate space point which has high reliability as seed. The details of the reconstruction model fit to the original object more. Meanwhile the number of reconstruction points in the model with less texture has been significantly increased and the holes of the texture model are obviously reduced. The validity of the improved algorithm is proven to be more effective and practical.

Key words:  PMVS algorithm, Universal-RANSAC, double constraints, multi-view stereo reconstruction

摘要： PMVS（patch-based multi-view stereo）算法以其良好的表现，在多视立体领域得到广泛应用。然而，算法存在重建模型细节丢失与重建点位置不够精确的问题，这种情况在输入图片较少，重建场景纹理不明显时尤为严重。针对这些不足，对去除候选误匹配点及对种子点置信度的排序进行了研究：引入USAC（Universal-RANSAC）去除候选误匹配点方法；提出双约束条件策略，筛选出候选空间点中置信度较高的点作为种子点。重建模型细节与原物体的契合度有了很大提高，纹理较少模型的重建点云数明显增加，漏洞也明显减少。通过在真实数据集上的实验，验证了改进算法具有更强的有效性和实用性。

关键词: PMVS算法, Universal-RANSAC, 双约束条件, 多视立体重建