Conceitos Básicos
Recovering relightable neural surfaces using pre-integrated rendering for advanced applications.
Resumo
This paper introduces NeuS-PIR, a method for reconstructing relightable neural surfaces from multi-view images or video. It utilizes implicit neural surface representation to factorize geometry, material, and illumination. The joint optimization addresses ambiguity in reconstruction, enabling relighting and material editing. Indirect illumination fields are distilled for complex lighting effects. Qualitative and quantitative experiments show superior performance over existing methods.
- Introduction to inverse rendering challenges in computer vision.
- NeuS-PIR methodology overview focusing on geometry, material, and illumination factorization.
- Comparison with existing methods like NeRFactor and NVDiffrec.
- Evaluation on synthetic datasets showcasing improved performance in relighting tasks.
- Real-world dataset evaluation demonstrating superior results in novel view synthesis.
- Ablation study highlighting the importance of SDF loss and regularization techniques.
- Training strategy analysis showing benefits of early introduction of the material module.
- Discussion on indirect illumination distillation and comparison with Neural-PIL's environment modeling approach.
Estatísticas
論文では、NeuS-PIRが既存の手法を上回ると示す定量的および定性的実験が行われました。
ネットワークアーキテクチャには、多層パーセプトロン(MLP)などが使用されています。
学習率は0.01であり、Adamオプティマイザーが使用されました。
Citações
"Our method enables advanced applications such as relighting."
"Qualitative and quantitative experiments have shown that NeuS-PIR outperforms existing methods."