SplatLoc: 3D Gaussian Splatting-based Visual Localization for Augmented Reality.

Area of Science:

• Computer Vision
• Computer Graphics
• Robotics

Background:

• Visual localization is crucial for augmented reality (AR) devices to determine their 6-DoF pose within pre-built maps for virtual content rendering.
• Existing methods often struggle with novel view rendering and demand substantial map storage capacity.

Purpose of the Study:

• To develop an efficient visual localization technique that supports high-quality rendering with minimal parameters.
• To address the limitations of existing approaches in terms of rendering capabilities and storage requirements.

Main Methods:

• Leveraging 3D Gaussian primitives for scene representation.
• Developing an unbiased 3D scene-specific descriptor decoder for Gaussian primitives to ensure accurate 2D-3D correspondences.
• Implementing a salient 3D landmark selection algorithm for efficient localization using a subset of primitives.
• Regularizing key Gaussian primitives to mitigate anisotropic effects and enhance localization accuracy.

Main Results:

• The proposed method demonstrates efficient visual localization with fewer parameters.
• Achieves high-quality rendering capabilities, overcoming limitations of prior techniques.
• Experimental results show superior or comparable performance to state-of-the-art implicit-based visual localization approaches on standard datasets.

Conclusions:

• The 3D Gaussian primitive-based approach offers an effective solution for visual localization in AR applications.
• The method provides a balance between localization accuracy, rendering quality, and storage efficiency.
• This work advances the field by enabling more practical and performant AR experiences.