Speaker: Akihiko Torii
Description: The 1st half of the tutorial deals with the visual place recognition problem. This task is formulated as given a query image of a particular street or a building, we seek to find one or at most a few images in the geotagged database depicting the same place. We focus on localizing “a single query testing image” by matching to “a large image database”. This part consists of four sub-parts.
1. A brief introduction of the visual place recognition. We briefly discuss about the differences between street-level and landmark visual recognition scenarios.
2. Reviews of basic image retrieval techniques. We explain the basic mechanism of bag-of-visual-words based image retrieval in detail. Then, we discuss several recent techniques which can be advantageous to the visual place recognition.
3. Advances in visual place recognition. We introduce several interesting approaches which tackle the challenges due to the existence of problematic objects in the urban scenes, e.g. repetitive structures, relatively quickly changing objects, distinctive but appears repeatedly. These problems are addressed by adopting several ideas: by more suitable image description, using geotag as supervision, and using structured relationships in database images.
4. Finally, we show some links to the standard benchmarking datasets of place and landmark recognition. Also, we discuss how to evaluate the performance of place recognition.