ICSD-NeRF: Independent Canonical Spaces for Enhanced Dynamic Scene Modeling in Neural Radiance Fields

초록

Novel view synthesis for dynamic scenes is a critical challenge in computer vision and computational imaging. Despite significant advancements, generating realistic images from monocularly captured dynamic scenes remains a complex task. Recent methods leveraging neural radiance fields and 3D Gaussian splatting in canonical spaces have made notable progress. However, these approaches often estimate both color and geometric information within a single space, limiting their effectiveness in handling large deformations of dynamic objects or significant color variations. To address these challenges, we propose ICSD-NeRF, which optimizes color and geometry in separate canonical spaces. Additionally, we introduce decision fields to effectively distinguish and optimize static and dynamic objects, enabling dynamic regions to be disentangled from the static background during training. To further enhance the representation of geometric structures in static regions, we employ an MLP to refine geometric features. We validate our approach on widely used dynamic scene novel view synthesis datasets, demonstrating that ICSD-NeRF outperforms existing methods by achieving higher rendering accuracy. Notably, our method achieves higher PSNR scores on benchmark datasets than current state-of-the-art techniques.

키워드