CT-Loc: Cross-domain visual localization with a channel-wise transformer

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Cross-domain visual localization is challenging due to photometric and geometric differences between synthetic and real images.
• Existing methods often show suboptimal performance in real-world scenarios without 3D spatial mapping.
• Accurate camera pose estimation is crucial for various applications like autonomous navigation and augmented reality.

Purpose of the Study:

• To develop a novel deep learning approach for robust cross-domain visual localization without requiring 3D spatial mapping.
• To improve the accuracy of camera position and orientation estimation from real images.
• To address the limitations of current methods in handling domain gaps.

Main Methods:

• A channel-wise transformer localization (CT-Loc) framework is proposed, inspired by human landmark recognition.
• The framework encodes salient features of task-relevant objects using attention weights and saliency maps.
• A large-scale dataset of a mechanical room interior was created for real-world evaluation.

Main Results:

• CT-Loc significantly reduces position and orientation errors compared to state-of-the-art models like BIM-PoseNet.
• The method effectively learns and utilizes relevant visual structural patterns (e.g., floors, doors) while ignoring irrelevant objects.
• Robust camera-pose regression localization results were achieved without prebuilt maps.

Conclusions:

• The proposed CT-Loc framework offers a robust solution for cross-domain visual localization.
• The attention-based approach enhances localization accuracy by focusing on critical scene features.
• This method advances camera pose estimation capabilities in real-world applications without relying on 3D models.