Enhanced outdoor visual localization using Py-Net voting segmentation approach.

Area of Science:

• Computer Vision
• Robotics
• Geospatial Analysis

Background:

• Camera relocalization is crucial for autonomous systems.
• Existing methods struggle with outdoor scene complexity and scale.
• Scene coordinate regression methods show limitations in large-scale environments.

Purpose of the Study:

• To propose Py-Net, a novel visual localization method for robust outdoor camera relocalization.
• To improve accuracy and efficiency in large-scale, complex outdoor environments.
• To address limitations of existing methods in handling repetitive structures and low-texture images.

Main Methods:

• Py-Net employs a voting segmentation approach with a main encoder featuring a Py-layer.
• The Py-layer utilizes pyramid convolution for multi-level feature extraction with reduced parameters.
• Coordinate attention and deep over-parameterized convolution modules enhance feature correction and robustness.

Main Results:

• Py-Net achieved lower distance and angle errors in multiple outdoor scenes compared to existing methods.
• The method effectively utilizes landmark segmentation and voting maps for precise 3D spatial relations.
• Py-Net demonstrated a 31.85% reduction in parameters and a smaller model size (170MB vs. 236MB) compared to VS-Net.

Conclusions:

• Py-Net offers a more efficient and accurate solution for outdoor camera relocalization.
• The proposed architecture effectively extracts scene information and corrects features for improved robustness.
• Py-Net presents a significant advancement in visual localization for large-scale environments.