Adaptive Range guided Multi-view Depth Estimation with Normal Ranking Loss

Yikang Ding, Zhenyang Li, Dihe Huang, Kai Zhang, Zhiheng Li, Wensen Feng; Proceedings of the Asian Conference on Computer Vision (ACCV), 2022, pp. 1892-1908

Deep learning algorithms for Multi-view Stereo (MVS) have surpassed traditional MVS methods in recent years, due to enhanced reconstruction quality and runtime. Learning-based methods, on the other side, continue to generate overly smoothed depths, resulting in poor reconstruction. In this paper, we aim to boost depth estimation (BDE) for MVS and present an approach for reconstructing high-quality point clouds with precise depth prediction. This method is termed as BDE-MVSNet. We present a non-linear technique that derives an adaptive depth range (ADR) from the estimated probability, motivated by distinctive differences in estimated probability between foreground and background pixels. ADR offers accurate estimation while processing same-resolution depth maps in only two stages since the depth range is well-adapted for each pixel. ADR also tends to decrease fuzzy boundaries via upsampling low-resolution depth maps between stages. Additionally, we provide a novel structure-guided normal ranking (SGNR) loss that imposes geometrical consistency in boundary areas by using the surface normal vector. Extensive experiments on DTU dataset, Tanks and Temples benchmark, and BlendedMVS dataset demonstrate that our method outperforms known methods and achieves state-of-the-art performance.