Adaptive Feature Enhanced Multi-View Stereo With Epipolar Line Information Aggregation

Despite the promising performance achieved by the learning-based multi-view stereo (MVS) methods, the commonly used feature extractors still struggle with the perspective transformation across different viewpoints. Furthermore, existing methods generally employ a "one-to-many" strategy, computing the correlations between the fixed reference image feature and multiple source image features, which limits the diversity of feature enhancement for the reference image. To address these issues, we propose a novel Epipolar Line Information Aggregati(EIA) method. Specifically, we present a feature enhancement layer (EIA-F) that utilizes the epipolar line information to enhance image features. EIA-F employs a "many-to-many" strategy, adaptively enhancing the reference-source feature pairs with diverse epipolar line information. Additionally, we propose a correlation enhancement module (EIA-C) to improve the robustness of correlations. Extensive experiments demonstrate that our method achieves state-of-the-art performance across multiple MVS benchmarks, particularly in terms of reconstruction integrity.