Scene recognition with bag of visual nouns and prepositions

John Stalbaum, Hee Won Chae, Jae Bok Song

Research output: Contribution to journalArticlepeer-review


The loop closure problem is central to topological simultaneous localization and mapping (SLAM); by associating features between distant portions of a trajectory, the odometry error that has accumulated between two observations can be eliminated and a more consistent map can be built. Bayesian pattern recognition techniques such as bag of visual words (BoVW) have recently shown outstanding results in solving the loop closure problem completely in image space using very simple, inexpensive cameras, without the requirement for highly accurate metric information, 3D reconstruction, or camera calibration. In this paper, a modified BoVW descriptor that incorporates simple geometric relationships within an image is used with the fast appearance-based mapping (FAB-MAP) algorithm. In direct comparisons with the traditional BoVW descriptor, an improved recall rate is observed with an acceptable increase in computational time. The proposal of a BoVW-compatible descriptor and the use of the proposed descriptor with a well-known BoVW classifier demonstrate the ability of the BoVW metaphor to be generalized, which could pave the way for more various BoVW descriptors in the same way that many individual visual feature descriptors exist within the computer vision community.

Original languageEnglish
Pages (from-to)115-125
Number of pages11
JournalIntelligent Service Robotics
Issue number2
Publication statusPublished - 2015 Apr 1

Bibliographical note

Publisher Copyright:
© 2015, Springer-Verlag Berlin Heidelberg.


  • Bag of visual words
  • Loop closure
  • Place recognition
  • SLAM
  • Scene recognition

ASJC Scopus subject areas

  • Computational Mechanics
  • Engineering (miscellaneous)
  • Mechanical Engineering
  • Artificial Intelligence


Dive into the research topics of 'Scene recognition with bag of visual nouns and prepositions'. Together they form a unique fingerprint.

Cite this