Reading #25: A Descriptor for Large Scale Image Retrieval Based on Sketched Feature Lines (Eitz)

Comments

Sam

Summary

Internet changed the way we use and retreive information, now the problema is that there is so much information that efficient queries are necessary. Searching engines like google have done an amazing job at searching text. However searching for images is not always easy, particularly when the images in the database do not contain metadata. The paper proposes a way of searching images based on sketches of the desired image. The problem here is that the processing in each image has to be simple enough to allow the query to return results in feasible time. The contribution is a descriptor based on structure tensors that allows easy and efficient implementation to obtain query results in low times over a large database (less than 4 seconds for a 1.5 Million images database).

Discussion

This application shows a really interesting way of doing queries. The results are very impressive in terms of time and accuracy. The resulting images are not always as expected (draw a building return a railroad) but this is not necessarily a bad thing, as actually it gives the user the feedback that he draw a building that looks like a railroad. I don’t think sketch queries will replace text queries for images, as in usual cases it is easier to search for “tree” than to draw a tree. However they are multiple cases where this can be a very nice option as when a very abstract picture is searched or the user is not really sure of how to describe the image he clearly has on his mind. Also this enables images in the result set that do not have the adequate metadata, and allows Multilanguage search. (If I search for “tree” the image will metadata “arból” will most likely be discarded even if they mean the same).