arXiv 2509.16909

SLAM-Former: Putting SLAM into One Transformer

By Yijun Yuan, Zhuoguang Chen, et al.

Published 2025-09-21

Wiki summary

Explore the paper's summary, context, and related research on Papiers.

We present SLAM-Former, a novel neural approach that integrates full SLAM capabilities into a single transformer. Similar to traditional SLAM systems, SLAM-Former comprises both a frontend and a backend that operate in tandem. The frontend processes sequential monocular images in real-time for incremental mapping and tracking, while the backend performs global refinement to ensure a geometrically consistent result.…

View the original paper on arXiv