Unsupervised Visual Landmark Extraction For Place Recognition

Sariyanidi E., Temeltaş H.

Conference on Full Motion Video (FMV) Workflows and Technologies for Intelligence, Surveillance, and Reconnaissance (ISR) and Situational Awareness, Maryland, United States Of America, 23 - 24 April 2012, vol.8386 identifier identifier

  • Publication Type: Conference Paper / Full Text
  • Volume: 8386
  • Doi Number: 10.1117/12.920680
  • City: Maryland
  • Country: United States Of America
  • Istanbul Technical University Affiliated: Yes


This paper presents an approach to detect visually salient patches in order to use them for visual place recognition. We formulate the saliency detection problem as an optimization problem, and define an energy function which describes the distinctiveness of a given image patch. We employ a Branch & Bound based search technique to efficiently find the global optimum of the energy function. Moreover, we use integral images to further increase the efficiency of the approach. The proposed saliency detection technique is able to detect patches which are suitable to be used as visual landmarks, and it performs with very high efficiency.