# BAMF-SLAM: Bundle Adjusted Multi-Fisheye Visual-Inertial SLAM Using Recurrent Field Transforms

**Authors:** Wei Zhang, Sen Wang, Xingliang Dong, Rongwei Guo, Norbert Haala  
**arXiv:** [2306.01173](https://arxiv.org/abs/2306.01173)  
**Date:** 2023-06-01  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2306.01173v2/

## Key points

- Authors introduce BAMF-SLAM, a cutting-edge multi-fisheye visual-inertial SLAM system
- System leverages Bundle Adjustment (BA) and recurrent field transforms (RFT) for precise state estimation
- Operates directly on raw fisheye images to maximize Field-of-View (FoV)
- Integrates multi-camera inputs and inertial measurements tightly through a unified factor graph
- Maintains global consistency by efficiently identifying loop closures using wide FoV of fisheye cameras
- Excels at wide baseline loop closing with minimal overlap due to broad convergence basin of RFT
- Introduces semi-pose-graph BA method to balance computational expense and accuracy effectively
- Validated performance on TUM-VI, Hilti-Oxford, and Newer College datasets showing superior results compared to existing methods
- VIO version secures second place in Hilti SLAM Challenge 2022; complete system surpasses winning approach in subsequent submissions

## Summary

In their paper titled "BAMF-SLAM: Bundle Adjusted Multi-Fisheye Visual-Inertial SLAM Using Recurrent Field Transforms," authors Wei Zhang, Sen Wang, Xingliang Dong, Rongwei Guo, and Norbert Haala introduce a cutting-edge multi-fisheye visual-inertial SLAM system that leverages Bundle Adjustment (BA) and recurrent field transforms (RFT) to achieve precise and robust state estimation in challenging environments. The system operates directly on raw fisheye images, maximizing the wide Field-of-View (FoV) offered by fisheye cameras. To address the low-texture challenge often encountered in such scenarios, the authors integrate multi-camera inputs and inertial measurements tightly through a unified factor graph. One of the key strengths of BAMF-SLAM is its ability to maintain global consistency by exploiting the wide FoV of fisheye cameras to identify potential loop closures efficiently. Powered by the broad convergence basin of RFT, the system excels at performing wide baseline loop closing with minimal overlap. Additionally, a semi-pose-graph BA method is introduced to avoid the computational expense associated with full global BA. By combining relative pose factors with loop closure factors, global states can be adjusted effectively with modest memory usage while preserving high accuracy. The performance of BAMF-SLAM is validated through evaluations on TUM-VI, Hilti-Oxford, and Newer College datasets, demonstrating superior results compared to existing methods. Notably,in the Hilti SLAM Challenge 2022,the VIO version of BAMF-SLAM secures second place.In subsequent submissions,the complete system incorporating the global BA backend surpasses even the winning approach. This innovative approach showcases significant advancements in multi-fisheye visual-inertial SLAM technology and holds promise for applications requiring accurate state estimation in complex scenarios.