https://nesf3d.github.io/

他们利用depth了吗？他们的precision如何？

Introduction

Summary

1. Feature Grid Fs; 2d segmentation as GT.

2. generate a semantic field s(x) of probability distributions over semantic categories

   

3. NeSF: how to generate 3d segmentation?

4. How to improve 2d segmentation? it is hard.

5. density field to semantic field. why not σ + color ⇒ semantic?

   1. color field is “faked” and no means to volume segmentation.

Vs Semantic-NeRF

1. Semantic-NeRF [149], regresses a per-3D point semantic class in addition to radiance and density.
   1. only applicable to novel views within the same scene and does not provide the form of generalization one expects in classical semantic segmentation: the ability to infer semantics on novel scenes.
2. Is it too simple? No. why oral.
   1. Label Propagation: from sparse point labels to dense labels.

Strengths

Weakness

Experiments

References