LGSDF: Continual Global Learning of Signed Distance Fields Aided by Local Updating

Yufeng Yue1, Yinan Deng1, Yujie Tang1, Jiahui Wang1, Yi Yang1,
1Beijing Institute of Technology
Paper arXiv Code

Abstract

Implicit reconstruction of ESDF (Euclidean Signed Distance Field) involves training a neural network to regress the signed distance from any point to the nearest obstacle, which has the advantages of lightweight storage and continuous querying. However, existing algorithms usually rely on conflicting raw observations as training data, resulting in poor map performance. In this paper, we propose LGSDF, an ESDF continual Global learning algorithm aided by Local updating. In the front-end, anchors are uniformly distributed throughout the scene and incrementally updated based on preprocessed sensor observations, reducing estimation errors caused by limited viewing directions. In the back-end, a randomly initialized implicit ESDF neural network undergoes continuous self-supervised learning, driven by strategically sampled anchors, to produce smooth and continuous maps. The results on multiple scenes show that LGSDF can construct more accurate SDFs and meshes compared with SOTA ESDF mapping algorithms.

Framework

teaser

LGSDF takes a sequence of posed depth images as input and continuously trains a randomly initialized neural network as the final implicit ESDF map. LGSDF consists of two main modules: the front-end and the back-end. The front-end Local Updating Module focuses on incrementally updating local anchors based on current observations. In contrast, the back-end Global Learning Module handles the curation of training data (i.e., anchors) and employs specific network optimization strategies. As the posed depth images are continuously input, the network gradually restores the full SDF of the observed scene.

Results

The Mapping Process in 10x

Tips: Videos play better in full screen!

Full ESDF

Image 1 Image 2 Image 3
3D Mesh
Our Slice
GT Slice

You can slide to get different Z 's ESDF slices :

Topmost Bottommost
Image 1 Image 2 Image 3
3D Mesh
Our Slice
GT Slice

You can slide to get different X 's ESDF slices :

Leftmost Rightmost
Image 1 Image 2 Image 3
3D Mesh
Our Slice
GT Slice

You can slide to get different Y 's ESDF slices :

Backmost Frontmost

2D ESDF Slices and Meshes Comparison

teaser

3D Reconstruction

Visualisation of scene reconstructed by different methods.


Voxblox

iSDF

LGSDF

GT

BibTeX

@article{lgsdf,
      title={LGSDF: Continual Global Learning of Signed Distance Fields Aided by Local Updating},
      author={Yue, Yufeng and Deng, Yinan and Tang, Yujie and Wang, Jiahui and Yang, Yi},
      journal={arXiv preprint arXiv:2404.05187},
      year={2024}
    }