September 21, 2010 @ 11:30am : Josef Sivic (INRIA & ENS)

Location : Warren Weaver Hall (WWH) 1302 (Directions)

Large scale visual place recognition with Google Street-View images

Josef Sivic (INRIA-Willow Project, Laboratoire d'Informatique de l'Ecole Normale Superieure)

We seek to recognize the place depicted in a query image using a database of ``street side” images annotated with geolocation information.

Efficient visual search for particular objects and places has demonstrated impressive retrieval results across differences in scale, viewpoint, and some amount of lighting variations. However, the image collection is typically treated as an unstructured, unordered set of images.

In this talk I will discuss a method, developed for large scale visual place recognition, which takes advantage of the structured nature of street-side imagery. In particular, we use the geotags available with each image as a form of supervision to offline modify the image database by detecting and suppressing image features, e.g. on trees or road-markings, which frequently occur in the database and hence cause significant confusion between different places. I will show the method combines well with the bag-of-features model for efficient visual search, including query expansion, and demonstrate significantly improved place recognition performance while reducing the database size.

Finally, I will discuss an application of particular place search to image inpainting and present a method to replace a user specified target region of a photograph by using registered photographs of the same place downloaded from the Internet. I will demonstrate removal of objects and people in challenging photographs of Oxford landmarks containing complex image structures.

Joint work with J. Knopp, T. Pajdla, O. Whyte and A. Zisserman

seminars/seminaritems/2010-09-21.txt · Last modified: 2011/02/27 17:28 by yann