Förståelse för kraven på bildåtervinning för visuell lokalisering

Visual Localization is a key task in autonomous systems and robotics, consisting of the estimation of the pose of a camera, given the image it is capturing. Despite different approaches exist, they always rely on comparing the query image to reference images with known poses to find correspondences between observed objects. To improve efficiency, only a subset of reference images are matched, determined by image retrieval pipelines. Image retrieval methods adopted in the literature are usually trained on different tasks with looser retrieval objectives, such as Visual Place Recognition. This work studies the impact of retrieval algorithms in Visual Localization, focusing on the specific requirements for the task. Firstly, the ideal requirements that image retrieval methods should meet are determined, replacing the retrieval stage with an image selection based on different criteria, such as high covisibility or spatial closeness; comparing localization results varying the selection criterion, insights on the ideal characteristics are drawn, while also setting loose upper bounds on Visual Localization performances for actual retrieval methods. Secondly, current retrieval methods are benchmarked, to determine the state of the art on the task, and to evaluate the performance gap with respect to the ideal bounds set. Lastly, small variations of existing retrieval methods are tested, such as re-ranking retrieved images, trying to determine the magnitude of future work needed to improve localization performances. The experiments highlight the importance of high covisibility between query and reference images, which is often related to their pose similarity. Furthermore, upper bounds correctly localize 20% of dataset images that actual retrievals are unable to, showing that the current models limit Visual Localization and tailoring is needed. The small pipeline variations tested reduce the performance gap, highlighting the viability of minor adjustments such as fine-tuning. Overall, this work shows the retrieval requirements for Visual Localization, setting the research directions to tailor models for the task to remove the retrieval bottleneck from the pipeline.