Attribute Analysis of Geo-tagged Images for City Perception

Bolei Zhou Liu Liu Aude Oliva Antonio Torralba
Massachusetts Institute of Technology



Download Paper

After hundreds of years of human settlement, each city has formed its own identity, distinguishing itself from other cities. When people travel in a city they will take photos of the city and upload them to social networks. These massive geo-tagged photos reflect the people's perception for the city and the identity of the city. In this work, we propose to characterize the identity of a city via an attribute analysis of 2 million geo-tagged images from 21 cities over 3 continents. We first build a higher-level set of 7 city attributes, tailored to the form and function of cities to represent the geo-tagged images. Then we design a machine learning algorithm to predict the city attributes for the geo-tagged images. By mapping the geo-tagged images predicted with city attributes, we analyze how people interact with the city and how the urban life is spatially distributed. We further conduct a city recognition experiments to identity images with salient city identity on each attributes. Our work shows the potential application of computer vision techniques to urban planning and urban computation.

Check out this page for a few results, and read our paper for full details.


City Attribute Analysis

We predicted the city attributes from scene attributes for geo-tagged images of several cities as follows:

  • Boston
  • New York
  • San Francisco
  • Paris
  • London
  • Tokyo
  • Demo of Attribute Prediction

    *Updated demo from Places-CNN with scene recognition and scene attribute prediction*

    Spatial Distribution of City Attributes

    After we predicted the city attributes of the images, we could plot the spatial distribution of city attributes on a map based on the geo-tags of the predicted images, called as perception map. Here we plot the perception maps of Bercelona and New York City as follows [More]

    City Identity Images

    We conducted city classification experiments and found the images with high city identity value on each city attribute below.

  • Green Space
  • Transportation
  • Architecture
  • Vertical Building
  • Water Coverage
  • Athletic Activity
  • Social Activity
  • City Similarity

    The similarity graph of cities estimated by our method is shown below. The nodes indicate cities, the thickness of the edges indicates the similarity of cities. Cities grouped together have the similar visual appearance.

    MIT City Database

    A geo-tagged image database called MIT City Database is available to download. MIT City Database contains around 2 million geo-tagged images from 21 cities. Some representative images of each city are shown below.

    [Download Database]

    Notice: Image URLs are cropped from Panoramio, which are for academic research and education purposes only.

    References

    B. Zhou, L. Liu, A. Oliva, A. Torralba. "Recognizing City Identity via Attribute Analysis of Geo-tagged Images." In Proceedings of 13th European Conference on Computer Vision (ECCV 2014).

    @article{zhou2014cityidentity,
      title={{Recognizing City Identity via Attribute Analysis of Geo-tagged Images}},
      author={Zhou, B. and Liu, Liu. and Oliva, A. and Torralba, A.},
      journal={ECCV},
      year={2014}
    }

    L. Liu, B. Zhou, J. Zhao, B. Ryan. "C-IMAGE: City Cognitive Mapping Through Geo-Tagged Photos." In Proceedings of the Association of Collegiate Schools of Planning (ACSP 2014).