Map-Free Visual Relocalization Using Relative Pose Regression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Visual relocalization in computer vision applications, such as augmented reality and robotics, is challenging due to the ambiguity of visual data, requiring pre-generated 3D maps and extensive scanning processes, which restricts the flexibility and speed of pose estimation.
Innovation Solution
A map-free approach using a relative pose regression network that determines the pose of a camera based on a single reference image and a query image, employing a Siamese network, correlation network, and multilayer perceptron to generate relative pose estimates without relying on explicit 2D correspondences, and incorporating depth prediction for improved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If pre-generated 3D maps are used for visual relocalization, then measurement precision is improved, but device complexity and preparation time increase
Solution Approach 1:
The patent extracts the essential relocalization functionality from the complex pre-generated 3D map framework. Instead of requiring full 3D maps, the system uses only a single reference image captured at a known pose, eliminating the need for extensive mapping infrastructure while maintaining pose estimation capability through learned feature correspondences between the reference image and query images
Solution Approach 2:
The patent creates a simplified copy of the mapping concept by using a single reference image instead of a complete 3D map. This reference image serves as a lightweight surrogate that contains sufficient information for relocalization when combined with deep learning features, drastically reducing the complexity of the mapping system while preserving the essential functionality
2Measurement precision
If extensive scanning processes are used to create 3D maps, then measurement precision is improved, but productivity and speed of deployment decrease
Solution Approach 1:
The patent performs preliminary action by capturing a single reference image at a known pose before deployment. This pre-captured image serves as the foundation for all subsequent relocalization operations, eliminating the need for extensive scanning processes during deployment while maintaining accuracy through the learned relationship between the reference image and query images
Solution Approach 2:
The patent skips the time-consuming 3D mapping and scanning processes entirely. By using a single reference image combined with deep learning feature extraction and relative pose regression, the system rushes through the deployment process instantaneously, achieving both high speed and maintained precision without traditional mapping overhead
3Reliability
If traditional feature matching with 3D maps is used, then reliability is improved, but device complexity increases
Solution Approach 1:
The patent substitutes the mechanical 3D map matching system with a learning-based approach. Instead of extracting features and matching them against pre-generated 3D maps, the system uses a Siamese network to extract features from both reference and query images simultaneously, followed by a correlation network and relative pose regression network to directly compute the pose, replacing complex mechanical mapping infrastructure with a streamlined neural network pipeline
Data Source
AI summary
A method or a system for map-free visual relocalization of a device. The system obtains a reference image of an environment captured by a reference pose. The system also receives a query image taken by a camera of the device. The system determines a relative pose of the camera of the device relative to the reference camera based in part on the reference image and the query image. The system determines a pose of the query camera in the environment based on the reference pose and the relative pose.