InLoc: Indoor Visual Localization with Dense Matching and View Synthesis

Area of Science:

• Computer Vision
• Robotics
• Geographic Information Systems

Background:

• Accurate 6-DoF pose estimation is crucial for indoor navigation and augmented reality.
• Existing visual localization methods struggle with large-scale indoor environments and weakly textured scenes.

Purpose of the Study:

• To develop a robust and scalable visual localization method for predicting the 6-DoF pose of query photographs within large indoor 3D maps.
• To introduce a new dataset for evaluating indoor localization performance in realistic scenarios.

Main Methods:

• A three-step approach: efficient candidate pose retrieval, dense matching for pose estimation, and virtual view synthesis for pose verification.
• Utilizing dense matching to overcome challenges posed by weakly textured indoor environments.
• Employing virtual view synthesis for robustness against viewpoint, layout, and occlusion changes.

Main Results:

• The proposed method significantly outperforms state-of-the-art indoor localization techniques.
• Demonstrated superior performance on a new, challenging dataset featuring mobile phone captures.
• The method shows scalability for large-scale indoor environments.

Conclusions:

• The developed visual localization method offers a significant advancement for indoor pose prediction.
• The publicly released dataset and code will facilitate further research in large-scale indoor localization.