Sparse-View 3D Reconstruction

Sparse-view 3D reconstruction studies how to infer a coherent 3D scene from only a few images, enabling novel-view rendering, measurement, and simulation when dense capture is impractical. Methods optimize neural radiance fields or 3D Gaussian splats per scene, or use feed-forward models that predict geometry, camera poses, and appearance directly. Depth and stereo cues, pretrained 3D foundation models, uncertainty estimates, and diffusion-generated views are increasingly used to constrain poorly observed regions and guide view selection. The central challenge is completing unseen areas without letting inaccurate priors or generated details contradict the evidence, while balancing geometric accuracy, visual quality, generalization, and speed.