iS-MAP: Neural Implicit Mapping and Positioning for Structural Environments

Haocheng Wang, Yanlong Cao*, Yejun Shou, Lingfeng Shen, Xiaoyao Wei, Zhijie Xu, Kai Ren

*Corresponding author for this work

Research output: Chapter in Book or Report/Conference proceedingConference Proceedingpeer-review

Abstract

This work presents iS-MAP, a neural implicit RGB-D SLAM approach based on multi-scale hybrid representation in structural environments. iS-MAP encodes the scene using an efficient hybrid feature representation, which combines a 3D hash grid and multi-scale 2D feature planes. This hybrid representation is then decoded into TSDF and RGB values, leading to robust reconstruction and multilevel detail understanding. Additionally, we introduce Manhattan matching loss and structural consistency loss to fully incorporate the prior constraints of structured planes and lines. Compared with only color and depth losses, our structured losses are capable of guiding network optimization at the semantic level, resulting in more reasonable scene regularization. Experimental results on synthetic and real-world scene datasets demonstrate that our approach performs either better or competitive to existing neural implicit RGB-D SLAM methods in mapping and tracking accuracy, and predicts the most plausible reconstruction results for the unobserved structural regions. The source code will be released soon.

Original languageEnglish
Title of host publicationComputer Vision – ACCV 2024 - 17th Asian Conference on Computer Vision, Proceedings
EditorsMinsu Cho, Ivan Laptev, Du Tran, Angela Yao, Hongbin Zha
PublisherSpringer Science and Business Media Deutschland GmbH
Pages367-383
Number of pages17
ISBN (Print)9789819609680
DOIs
Publication statusPublished - 2025
Event17th Asian Conference on Computer Vision, ACCV 2024 - Hanoi, Viet Nam
Duration: 8 Dec 202412 Dec 2024

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume15480 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference17th Asian Conference on Computer Vision, ACCV 2024
Country/TerritoryViet Nam
CityHanoi
Period8/12/2412/12/24

Keywords

  • Neural implicit mapping
  • RGB-D SLAM
  • Self localization
  • Structural constraints

Fingerprint

Dive into the research topics of 'iS-MAP: Neural Implicit Mapping and Positioning for Structural Environments'. Together they form a unique fingerprint.

Cite this