基于SAM分割模型的多视图立体匹配算法OA
Multi View Stereo Matching Algorithm Based on SAM Segmentation Model
为了提高弱纹理区域、光度变化等复杂情况下多视图三维重建算法的完整性,文中提出了基于SAM(Seg-ment Anything Model)分割模型的平面先验辅助多视图立体匹配算法.对输入的多视图图像进行分割,分析分割区域的特征点分布规律,提取弱纹理区后生成先验图.采用空间传播与随机传播相结合的方式到初始深度图,滤波后根据先验图得到先验深度图.将初始深度图与先验深度图融合,使用多视图几何一致性约束生成最终深度图.在ETH3D和Tanks and Temples数据集进行实验.结果表明,相较于OpenMVS(Multi-View Stereo reconstruction library)算法,所提算法在5 cm和10 cm的误差下的F1分数分别提高了5.26%,3.32%,完整性分别提高了7.15%,5.29%,准确性分别提高了2.58%,0.99%.所提算法对完整性要求较高、大尺度场景平面较多的三维重建具有一定优势.
To improve the integrity of multi-view 3D reconstruction algorithms in complex situations such as weak texture regions and photometric variations,a planar prior assisted multi-view stereo matching algorithm based on the SAM(Segment Anything Model)segmentation model is proposed.The input multi-view image is segmented,the distribution law of feature points in the segmented area is analyzed,and the weak texture area is extracted to gen-erate the prior image.The initial depth map is obtained by combining spatial propagation with random propagation.After filtering,the prior depth map is obtained based on the prior map.The initial depth map is fused with the prior depth map,and the final depth map is generated using the multi-view geometric consistency constraint.Experi-ments are conducted on the ETH3D and Tanks and Temples datasets.The results show that compared with the OpenMVS(Multi-View Stereo reconstruction library)algorithm,the F1 scores of the proposed algorithm under er-rors of 5 cm and 10 cm have increased by 5.26%and 3.32%respectively,the completeness has increased by 7.15%and 5.29%respectively,and the accuracy has increased by 2.58%and 0.99%respectively.The proposed algorithm has certain advantages for 3D reconstruction with high integrity requirements and a large number of large-scale scene planes.
汪家通;林勇;吴添民
上海理工大学 健康科学与工程学院,上海 200093上海理工大学 健康科学与工程学院,上海 200093上海理工大学 健康科学与工程学院,上海 200093
信息技术与安全科学
SAM分割模型平面先验多视图立体匹配三维重建图像分割几何一致性先验深度图
SAM segmentation modelplane priormulti-view stereo matching3D reconstructionimage segmentationgeometric consistencyprior depth map
《电子科技》 2026 (10)
24-33,10
国家自然科学基金(81801797) National Natural Science Foundation of China(81801797)
评论