A list of visual(camera) re-localization researches. Visual relocalization refers to the problem of estimating the position and orientation using the information of an existing prior map based on the image captured from visual sensors. We sort out these methods according to the type of map, mainly focusing on image databases and point cloud maps.
This document will be continuously updated and there may inevitably be errors and omissions. Please feel free to point them out through issues or pull requests.
- Awesome-camera-relocalization-in-prior-map
The image database map consists of a series of images with information. The images in the image database can be captured by cameras with different intrinsic parameters. The poses of images in image databases are usually obtained by the SfM algorithm or external measurement devices such as GPS.
The core idea of image retrieval-based methods is to first search for the reference image that is most similar to the query image from the image database. Then, based on the relative pose relationship between the query image and the reference image, as well as the pose label of the reference image, the pose information of the query image is calculated.
The pose regression method extracts high-dimensional features of the query image through an end-to-end deep neural network, and then uses the high-dimensional features to directly regress the camera pose. We classify them based on different input.
Point cloud maps consist of a series of unordered points with three-dimensional coordinates in 3D space. Since the query image and the point cloud map are of different modalities, they cannot be directly compared. The key issue is how to establish data association. Existing methods are divided into three categories：(1) Feature based method(Visual Point Cloud): match by image features; (2): 3D
