MrNeRF (@janusch_patas)
2024-11-01 | โค๏ธ 136 | ๐ 19
[NeurIPS โ24] PhyRecon: Physically Plausible Neural Scene Reconstruction
TL;DR: PhyRecon harnesses both differentiable rendering and differentiable physics simulation to achieve physically plausible scene reconstruction from multi-view images.
Contributions (cited):
-
We introduce the first method that seamlessly bridges neural scene reconstruction and physics simulation through a differentiable particle-based physical simulator and the proposed SP-MC that efficiently transforms implicit representations into explicit surface points. Our method enables differentiable optimization with both rendering and physical losses
-
We propose a novel method that jointly models rendering and physical uncertainties for 3D reconstruction. By dynamically adjusting the per-pixel rendering loss and physics-guided pixel sampling, our model significantly improves the reconstruction of thin structures.
-
Extensive experiments demonstrate that our model significantly enhances reconstruction quality and physical plausibility, outperforming state-of-the-art methods. Our results exhibit substantial stability improvements, signaling broader potential for physics-demanding applications.
๋ฏธ๋์ด
