🚨 Are neural implicit representations applicable for larger-scale SLAM? Check out our NICE-SLAM👍! #CVPR2022
NICE website: pengsongyou.github.io/nice-slam
NICE code: github.com/cvg/nice-slam
NICE collaborations w/ Zihan Zhu (undergrad) et al.
1/6
Conversation
Replying to
The key idea of NICE-SLAM is super simple 😎:
➡️ Use the hierarchical feature grids as the scene representation
➡️ Incorporate the inductive bias of pretrained tiny occupancy MLPs
➡️ Backpropagate the NeRF-like volume rendering loss for mapping and tracking, alternatively
2/6
2
1
14
Compared to the seminal work iMAP that uses a single MLP as the scene representation, NICE-SLAM can significantly improve the quality of both mapping and tracking, as shown below.
Moreover, using hierarchical feature grids guarantees fast convergence & much less runtime.
3/6
1
10
To further demonstrate the scalability of NICE-SLAM, we also capture a sequence of an apartment with multiple rooms with a Kinect Azure.
With such a large scene, NICE-SLAM still works well because we can update locally the feature grids, unlike the global update of iMAP.
4/6
1
7
Finally, special thanks goes to iMAP authors , who greatly inspired us by demonstrating the possibility of implicit SLAM via continual training of an MLP!
We also released our re-implementation of iMAP at github.com/cvg/nice-slam#
5/6
2
16
That is all, we hope you will love NICE-SLAM as much as we do!
We're excited to tell you more about how NICE it is in New Orleans #CVPR2022 😁
6/6
read image description
ALT
3
18
