Three-Dimensional Reconstruction Optimization Algorithm Under Large Viewpoint Variations Scenes

To address the challenge of handling the geometric and appearance changes caused by large viewpoint variations in 3D reconstruction, we propose a deep learning-based optimization approach. Traditional SfM methods struggle with accurate reconstruction due to significant transformations in background, shape, scale, and texture caused by these variations. In our approach, we first utilize a deep learning-based SuperPoint feature extraction algorithm and SuperGlue feature matching algorithm. The match matrix is then verified and filtered using geometric consistency. Additionally, we employ feature extension techniques to enhance the descriptive capability of features under different viewpoints, thereby improving the model's robustness and generalization ability. To further enhance accuracy and stability, we apply geometric optimization through BA. Finally, we perform an incremental 3D reconstruction to select the best model. Comparing our optimization framework with traditional and deep learning-based baselines, we demonstrate its ability to produce more accurate 3D models in scenarios with large viewpoint variations, as shown on the Haiper dataset.