Real-World Evaluation of Monocular Visual SLAM:ORB-SLAM3, DROID-SLAM, DPVO, and DPV-SLAM

In this paper, we evaluate monocular visual SLAM methods in real-world environments. The methods for eval- uation are ORB-SLAM3, DROID-SLAM, DPVO, and DPV- SLAM. Additionally, we incorporate our proposed method that improves DPV-SLAM using AnyLoc loop closure. ORB-SLAM3 is a classical feature-based visual SLAM method. DROID-SLAM, DPVO, and DPV-SLAM are learning-based methods. DPV-SLAM is the improved method of DPVO, and it has loop closure capa- bility. However, conventional loop closure is not learning-based. Our AnyLoc loop closure utilizes learning-based loop detection, which outperforms classical Bag of Visual Words (BoVW) loop closure. Experimental results demonstrated the properties of each method. Our proposed method achieved high accuracy for SLAM in all environments and reasonable processing time.

In IEEE International Conference on Industrial Technology (ICIT 2026)