DUE-MVSNet: Learning Multi-view Stereo Based on Dual Uncertainty Estimation

Mingwei Cao, Siqi Nian, Jun Yi

Abstract: Many multi-view stereo networks employing a cascade structure can efficiently estimate depth while conserving memory. However, the accuracy of the depth map in the fine stage heavily relies on the depth map estimated in the coarse stage. Additionally, the multi-stage depth maps generated by the cascaded structure are simply used to calculate losses without further utilization, which results in the loss of inter-stage differentiation information. To address these issues, we propose a method called DUE-MVSNet. We use dual uncertainty estimation to mitigate the adverse effects introduced by the cascade structure. Specifically, we design an adjacent stage uncertainty (ASU) module and a pair-wise stage uncertainty (PSU) module. The ASU module dynamically adjusts the depth hypothesis range for the current stage by learning the uncertainty of the previous stage. The PSU module estimates the uncertainty between each pair of stages, mitigating the adverse effects of high uncertainty regions. We evaluated our method on the DTU and Tanks & Temples datasets. Experimental results show that our method achieves superior reconstruction results compared to the state-of-the-art methods. Code is available at https://github.com/caomw/due-mvsnet.

External IDs:doi:10.1007/978-3-031-82021-2_12