Abstract:
The crux of multi-view stereo reconstruction lies in the process of feature matching. To address the challenge of suboptimal feature extraction encountered by current mul...Show MoreMetadata
Abstract:
The crux of multi-view stereo reconstruction lies in the process of feature matching. To address the challenge of suboptimal feature extraction encountered by current multi-view stereo methods when dealing with regions characterized by weak texture or non-Lambertian surfaces, resulting in unsatisfactory reconstruction outcomes, this paper introduces a novel multi-view stereo network, AC-MVSNet. This network is founded on adaptive aggregation and coordinate attention mechanisms. This method firstly introduces an adaptive aggregation module, employing deformable convolution to dynamically tailor the size and shape of the receptive field based on features of varying scales, thereby obtaining a feature representation that encompasses both global and detailed information. Subsequently, following the adaptive aggregation module, a coordinate attention module is incorporated, utilizing coordinate information to dynamically adjust the weight of channel features, thereby enhancing the network's perception of detailed features such as boundary and texture information. Experimental results on the DTU dataset demonstrate that AC-MVSNet outperforms other learning-based methods in terms of its reconstruction performance.
Published in: 2024 IEEE 2nd International Conference on Control, Electronics and Computer Technology (ICCECT)
Date of Conference: 26-28 April 2024
Date Added to IEEE Xplore: 07 June 2024
ISBN Information: