Dynamic Geometry-Aware Depth Hypotheses and Edge-Guided Depth Refinement for Aerial Multi-View Stereo

Most of the learning-based methods for aerial multiview stereo (MVS) do not take into account the large depth variations and only rely on coarse stage depth information and uniform sampling to obtain the depth hypotheses, resulting in inaccurate depth estimation. In addition, aerial images cover a large area, and the overlap between multiple views is low, resulting in ambiguous matching in weak texture areas. To address these issues, the DGE-MVSNet is proposed, a novel coarse-to-fine network including a dynamic geometry-aware depth hypotheses (DGDHs) module and edge-guided depth refinement (EGDR) module. In detail, to generate reliable depth hypotheses, the DGDH module creates a neighborhood set for the pixel in the reference view, within which any point should be ensured to consistently lie on the same surface. Relying on this neighborhood set, the initial normal depth hypotheses can be estimated from neighboring points with the same normal. To further ensure accurate depth estimation, the EGDR module utilizes edge feature volumes to guide cost volume aggregation. The model uses cosine similarity to calculate view weights and generate invalidation masks to effectively filter out mismatched regions between source and reference views. During depth refinement, the initial depth is fused with the edge cost volume and combined with the attention block to improve depth estimation accuracy. Extensive experiments on the WHU-MVS and LuoJia-MVS datasets demonstrate that our model significantly outperforms the state-of-the-art methods. Compared to RED-Net, our model reduces the mean absolute error (MAE) by over 24.0%, and by more than 16.0% compared to Cas-MVSNet. To further validate the generalization capability of DGE-MVSNet, the evaluation is conducted on the WHU-OMVS dataset, specifically designed for city 3-D scene reconstruction. Results show that our method achieves a substantial improvement over the state-of-the-art AggrMVS, improving MAE by 29.2%.