Visual Localization via Virtual Views and Projective Transformations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current visual localization methods for indoor environments face challenges such as limited accuracy, perspective distortion, and the need for extensive reference images, making them inefficient and prone to errors, especially in texture-poor environments with complex geometric variations.
Innovation Solution
A method using pre-computed virtual views generated from sparse reference imagery, leveraging planar regions and projective transformations to create robust localization, allowing for accurate position and orientation determination using content-based image retrieval techniques, even with minimal reference images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If Bag-of-Features based image representations are used for localization, then robust description of the scene is achieved, but a huge amount of reference images are required
Solution Approach 1:
The patent creates virtual copies of reference images by applying projective transformations to generate synthetic views from sparse reference imagery. Instead of capturing every possible viewpoint with physical cameras, the system generates virtual images that simulate appearances from different locations and orientations, thereby reducing the need for extensive physical reference image collection while maintaining robust scene description capabilities
2Extent of automation
If reference images are captured along a single trajectory for mapping, then automated mapping is achieved, but the resolution of position and orientation estimates is drastically limited
Solution Approach 1:
The patent transitions from a single-trajectory 1D mapping approach to a multi-dimensional virtual viewpoint generation system. By creating virtual images at multiple locations and orientations throughout the 3D space, the system effectively adds spatial dimensions to the reference data, enabling high-resolution position and orientation estimates without requiring physically complex multi-trajectory mapping operations
3Reliability
If feature descriptors are made robust under perspective distortion, then recall is improved, but precision is reduced due to loss of distinctiveness
Solution Approach 1:
The patent changes the fundamental parameter of viewpoint representation by generating virtual images with known camera poses and positions. Instead of relying on perspective-distortion-robust feature descriptors that lose distinctiveness, the system uses projective geometry to accurately transform reference images to match query image perspectives, preserving feature distinctiveness while achieving perspective robustness through geometric transformation
Data Source
AI summary
In an embodiment of the invention there is provided a method of visual localization, comprising: generating a plurality of virtual views, wherein each of the virtual views is associated with a location; obtaining a query image; determining the location where the query image was obtained on the basis of a comparison of the query image with said virtual views.