SparsePose–NeRF: Robust Reconstruction Under Limited Observations and Uncalibrated Poses

Neural Radiance Fields (NeRF) reconstruction faces significant challenges under non-ideal conditions, such as sparse viewpoints or missing camera pose information. Existing approaches frequently assume accurate camera poses and validate their effectiveness on standard datasets, which restricts their applicability in real-world scenarios. To tackle the challenge of sparse viewpoints and the inability of Structure-from-Motion (SfM) to accurately estimate camera poses, we propose a novel approach. Our method replaces SfM with the MASt3R-SfM algorithm to robustly compute camera poses and generate dense point clouds, which serve as depth–space constraints for NeRF reconstruction, mitigating geometric information loss caused by limited viewpoints. Additionally, we introduce a high-frequency annealing encoding strategy to prevent network overfitting and employ a depth loss function leveraging Pearson correlation coefficients to extract low-frequency information from images. Experimental results demonstrate that our approach achieves high-quality NeRF reconstruction under conditions of sparse viewpoints and missing camera poses while being better suited for real-world applications. Its effectiveness has been validated on the Real Forward-Facing dataset and in real-world scenarios.