[Registration] global feature-extracting strategy

Reference

GaussReg: Fast 3D Registration with Gaussian Splatting

nerf2nerf: Pairwise Registration of Neural Radiance Fields

DReg-NeRF: Deep Registration for Neural Radiance Fields

NeRFuser: Large-Scale Scene Representation by NeRF Fusion

CL-NeRF: Continual Learning of Neural Radiance Fields for Evolving Scene Representation

• NeRF2NeRF [14] utilizes human-annotated key points to obtain an initial transformation and refines it using a surface field distilled from a pre-trained NeRF.

• DReg-NeRF [6] extracts features from the occupancy grid of NeRF and applies a decoupling model [38] for NeRF registration, eliminating the need for human interaction in the registration process. However, it’s hard to generalize to larger scenes due to its global feature-extracting strategy.
  • DReg-NeRF는 occupancy grid에서 특징을 추출하여 장면 간의 정합(registration)을 수행합니다. 이 과정에서:
    • 장면 전체의 특징을 한 번에 처리하려는 경향(global approach)이 있습니다.
    • 이는 장면 크기가 클수록 계산 비용이 증가하고, 복잡한 장면에서 국소적인 세부 정보를 놓칠 가능성이 있습니다.
    • 글로벌 접근법은 전역적인 요약 정보를 기반으로 하므로, 장면이 커질수록 세밀한 국소적 불일치(local mismatch)를 포착하는 데 한계가 생길 수 있습니다.

• NeRFuser [10] directly uses the structure from motion method to estimate the transformation using rendered images from NeRF which is very time-consuming.

• CL-NeRF [36] concentrates on the continual learning of NeRF models and proposes an expert adaptor for learning newly changed scenes without finetuning the whole network.