Are Large-Scale 3D Models Really Necessary for Accurate Visual Localization?

Area of Science:

• Computer Vision
• Robotics
• Autonomous Navigation

Background:

• Accurate visual localization is crucial for autonomous systems.
• 3D structure-based methods provide high accuracy but require extensive 3D model construction.
• 2D image retrieval methods are easier to maintain but traditionally less accurate.

Purpose of the Study:

• To investigate if large-scale 3D models are essential for accurate visual localization.
• To compare the accuracy of image-based methods with local reconstructions against 3D structure-based methods.
• To explore methods for optimizing the performance of image-based localization techniques.

Main Methods:

• Experimental evaluation on a large, challenging urban dataset with created reference poses.
• Combining 2D image retrieval with local 3D scene reconstruction.
• Analyzing pose accuracy and runtime costs of different localization approaches.
• Investigating the impact of utilizing known database image poses.

Main Results:

• Accurate visual localization is achievable without large-scale 3D models.
• Image-based methods combined with local reconstructions outperform state-of-the-art structure-based methods in pose accuracy.
• Higher runtime costs associated with combined methods can be mitigated by leveraging database image poses.
• The study demonstrates the potential of local models over extensive 3D models.

Conclusions:

• Large-scale 3D models may not be strictly necessary for accurate visual localization.
• Image-based methods coupled with local reconstructions offer a promising alternative for precise camera pose estimation.
• Further research is needed to accelerate local reconstruction processes for practical applications.