Retrieval: aim at ﬁnding as many relevant database images as possible for a given query image.
Localization: retrieve just one relevant image if the correspondences found by matching query and database image sufﬁce to estimate the camera pose of the query image.
- Recall-improving techniques such as query expansion do not improve the localization
fast, approximate spatial veriﬁcation
approaches could not be used
- for mobile App, efficiency is more important than accuracy – drop number of candidate images which could be localized
correspondences for camera pose estimation
- use image retrieval to find the similar views in dataset