EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Revisiting Keyframe Selection in Learning-Based Dense Monocular SLAM

Jaemin Choi, Juhwan Bae, Yun Gu Lee

IEEE Access · 2026 · 14 · 42342-42352

DOI: 10.1109/access.2026.3674511 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Jaemin Choi und 2 weiteren Autor:innen. Die Arbeit nennt 43 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Research Grant of Kwangwoon University, in 2025.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Recent learning-based dense monocular SLAM systems, such as VGGT-SLAM and MASt3R-SLAM, leverage powerful feed-forward 3D reconstruction models and achieve state-of-the-art performance even in uncalibrated monocular settings. However, in these systems, keyframe selection is often handled by simple frame-level heuristics based on low-level 2D signals, and the impact of input frame composition on feed-forward reconstruction and overall SLAM performance has not been systematically analyzed. In this paper, we reformulate keyframe selection in learning-based dense monocular SLAM as a problem of configuring the input frame set for feed-forward 3D reconstruction, rather than selecting individual frames. From this perspective, we propose a simple segment-level keyframe sampling strategy that aims to stabilize the composition of input frames. We analyze how this input configuration affects reconstruction stability and camera trajectory estimation. Experiments on public benchmarks demonstrate that the proposed strategy consistently improves camera trajectory RMSE compared to the original VGGT-SLAM and MASt3R-SLAM pipelines. These results highlight the importance of input frame configuration in learning-based SLAM and suggest a new perspective on keyframe selection beyond conventional frame-level decisions.

Förderung

  • Research Grant of Kwangwoon University, in 2025

Zitieren

APA

Choi, J., Bae, J., & Lee, Y. G. (2026). Revisiting Keyframe Selection in Learning-Based Dense Monocular SLAM. IEEE Access, 14, 42342-42352. https://doi.org/10.1109/access.2026.3674511

Vancouver

Choi J, Bae J, Lee YG. Revisiting Keyframe Selection in Learning-Based Dense Monocular SLAM. IEEE Access. 2026;14:42342-42352. doi:10.1109/access.2026.3674511

BibTeX
@article{Choi2026,
  author = {Choi, Jaemin and Bae, Juhwan and Lee, Yun Gu},
  title = {Revisiting Keyframe Selection in Learning-Based Dense Monocular SLAM},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {42342--42352},
  doi = {10.1109/access.2026.3674511}
}
Kostenlos starten →