Robust Neural Radiance Field Representation for Extrapolating Novel Views (RapNeRF)

Neural Radiance Fields (NeRF) have become a key technique for novel‑view synthesis, but their performance degrades when the test view is far from the training viewpoints. RapNeRF (RAy Priors) addresses this limitation by introducing two priors—Random Ray Casting (RRC) and Ray Atlas (RA)—to improve robustness for large view extrapolation.

Background : Traditional graphics pipelines achieve photo‑realistic rendering but are costly. Neural rendering leverages neural networks to encode geometry, material, and lighting from multi‑view images, enabling high‑quality 2D renderings. However, standard NeRF requires many images and struggles with sparse or unseen viewpoints.

Method :

Random Ray Casting (RRC) : For each training ray ov intersecting the surface at point v , a local polar coordinate system is built. The ray direction is perturbed within a limited angle (<30°) to generate a random ray ov' , providing on‑the‑fly data augmentation without pre‑computing extra images.

Ray Atlas (RA) : Surface points observed from multiple views are associated with view‑specific feature maps R(I_i) . These features are stored on the 3D surface, allowing reliable view‑dependent encoding during inference, which mitigates color bias and preserves fine details.

Training : RapNeRF is trained in two stages. First, a short‑epoch NeRF is trained to obtain coarse geometry. Then RRC and RA are activated (probabilities 0.7 and 0.5) for fine‑tuning, using both RGB loss and an opacity loss.

Experiments : On both synthetic and real datasets, RapNeRF outperforms state‑of‑the‑art baselines in PSNR, SSIM, and LPIPS, especially when test viewpoints are far from training views. Ablation studies show that RRC eliminates color shift while RA preserves reconstruction details.

Conclusion : RapNeRF provides a robust solution for 3D reconstruction in e‑commerce scenarios, enabling high‑quality novel‑view synthesis with sparse image inputs. Future work will explore further sparse‑image NeRF reconstruction.