DDL-MVS: Depth Discontinuity Learning for Multi-View Stereo Networks

We propose an enhancement module called depth discontinuity learning (DDL) for learning-based multi-view stereo (MVS) methods. Traditional methods are known their accuracy but struggle with completeness. While recent have improved completeness at the cost of accuracy, our DDL approach aims to improve while retaining in reconstruction process. To achieve this, we introduce joint estimation and boundary maps, where maps explicitly utilized further refinement maps. validate idea by integrating it into existing MVS pipeline depends on high-quality map estimation. Extensive experiments various datasets, namely DTU, ETH3D, “Tanks Temples”, BlendedMVS, show that method improves quality compared baseline, Patchmatchnet. Our ablation study demonstrates incorporating proposed significantly reduces error, instance, more than 30% DTU dataset, leads both smooth regions. Additionally, qualitative analysis has shown reconstructed point cloud exhibits enhanced without any significant compromise Finally, reveal model strategies exhibit strong generalization capabilities across datasets.