Multi-view Image Feature Alignment Strategy for Complex Scenes
Mengyin Ma, Qixin Su, Haotian Zhang, Xiaonan Song · 2024
Image feature alignment is a fundamental problem in the field of computer vision and is widely applied in various tasks such as 3D reconstruction, object tracking, and augmented reality. However, in complex real-world scenarios, challenging factors such as occlusion, deformation, and viewpoint variations pose significant challenges to the accuracy and robustness of feature alignment. To address this issue, this study proposes an innovative multi-view hierarchical feature alignment framework. The core of this framework lies in fully exploiting multi-modal heterogeneous information and fusing different cues, aiming to overcome the limitations of single visual cues while ensuring high-precision matching and significantly enhancing the algorithm's robustness. The framework consists of three innovative modules: multi-scale multi-view feature extraction, deep feature fusion based on attention mechanisms, and feature alignment based on multi-view geometric constraints. Comprehensive evaluations on three representative and challenging public datasets demonstrate that the proposed multi-view alignment algorithm achieves optimal accuracy and success rates. In complex scenarios with severe viewpoint changes, deformations, and occlusions, the performance advantage is even more pronounced.