Abstract:

Three-dimensional scene reconstruction from a set of two-dimensional images remains a significant and challenging research task in computer vision. Existing algorithms struggle to effectively model the relationships between images captured at different times， resulting in poor rendering quality for large-scale environments. To address this challenge， this paper proposes a neural surfels fusion splatting algorithm for real-time scene reconstruction. The algorithm progressively builds and updates scene models from streaming input data， incorporating neural surfel representations based on 2D Gaussian primitives and a fusion mechanism based on Gated Recurrent Units （GRU）. First， the algorithm achieves perspective-correct rendering by integrating Gaussian surfel splatting into the rendering pipeline and employing explicit ray-splat intersections， thereby enhancing depth consistency and geometric fidelity across multiple viewpoints. Then， to further improve the quality of reconstructed surfaces， the method combines depth pulling and normal consistency regularization， promoting smoother surface reconstruction and facilitating high-quality mesh extraction. The algorithm's performance was evaluated on two large-scale indoor/outdoor scene reconstruction datasets： DeepBlending and Tanks&Temples. Extensive experimental results demonstrate that the algorithm outperforms existing methods in both quantitative and qualitative assessments.

Key words: pose estimation, feature alignment, sequence information modeling, CNN