A multi-level composite attention-guided network for indoor visual localization

Accurate and robust camera pose estimation is essential for autonomous navigation and path planning in unmanned systems. To improve the localization accuracy in complex indoor scenes and mitigate information loss during feature extraction, we propose a multi-level composite attention-guided scene coordinate regression method. The proposed model predicts the mapping between 2D pixel points and 3D scene coordinates from a single RGB image. First, we introduce a Multi-level Feature Fusion Module (MFF), which employs global pooling and parallel branches to consolidate multi-level features, enhancing discrimination in repetitive structures and low-texture regions. Next, we design an Embedded Attention Module (EAM) to dynamically fuse multi-level features through parallel channel and spatial attention mechanisms, preserving edge details and suppressing noise. Finally, a differentiable random sample consensus algorithm is used to achieve robust fitting of pose parameters. Evaluation and analysis on common indoor public datasets demonstrate that the proposed method significantly improves localization performance. Additionally, extensive ablation evaluations confirm the effectiveness of the proposed Embedded Attention Module and Multi-level Feature Fusion Module in enhancing localization accuracy.