1: \begin{abstract}

2: % %paradigm

3: % We present FMapping, a robust and efficient neural field mapping framework that can continuously estimate a dense RGB point cloud map from an incomplete RGB stream for real-time dense RGB SLAM systems. 

4: % Our method eliminates the need for any geometric priors, such as pre-trained models and depth supervision.

5: 

6: % % When an RGB camera traverses the scene, 

7: % Specifically, 

8: % FMapping implicitly represents the scene as a 4D continuous radiance field, which is first factorized into a smaller number of low-rank terms to filter out the noise and retain only the most representative components. 

9: 

10: % It then exploits the coherent geometric cues to ensure proper sampling around shared intervals across neighboring frames for speedy convergence, during the system initialization stage. 

11: 

12: % The fully initialized FMapping gives a highly accurate initial trajectory and stabilized sample point distribution ready for continuously processing tracked new frames. 

13: 

14: % Notably, buttressed by such a latent space decomposition, FMapping exhibits high computational efficiency that is comparable to existing RGB-D solutions. Our method entails much lower uncertainty since the sample distribution is strongly skewed around the surface. 

15: 

16: % FMapping achieves it by estimating the point cloud distribution in the lower-dimensional feature space, effectively mitigating the uncertainty of the point sampling along the posed camera rays. 

17: 

18: % Extensive experiments demonstrate a siginicant boost in reconstruction fidelity compared with the previous neural implicit mapping methods for SLAM. Code is available at: \url{http://}

19: % % Please visit the project page for more details:

20: 

21: % \end{abstract}

22: