Frequently Asked Questions
1) What is StreetScore?
StreetScore is a machine learning algorithm that predicts how safe the image of a street looks to a human observer. We trained StreetScore to predict perceived safety using a 'training dataset' consisting of 3,000 street views from New York and Boston and their rankings for perceived safety obtained from Place Pulse — a crowdsourced survey. First, we converted the images in the training dataset to image features (or attributes), which represent the image's textures, colors and shapes. We then used machine learning to associate image features with scores of perceived safety collected from Place Pulse. To predict the perceived safety of a new image, StreetScore decomposes this image into features and assign the image a score based on the associations between features and scores learned from the training dataset.
2) Why did you build StreetScore?
The most important reason to build StreetScore is that it allows us to scale up the evaluation of street views for perceived safety by several orders of magnitude when compared to a crowdsourced survey. For instance, the map of New York City we are making available here contains more than 300,000 data points of perceived safety. Scoring 300,000 images using a crowdsourced visual survey would be impractical, since it would require obtaining more than 2 million pairwise comparisons of images. By using StreetScore, we are able to create high resolution maps that require less than 3,000 images (and only 24,000 comparisons) in our training dataset (certainly we could improve the quality of these maps by increasing the size of the training dataset).
3) How is Computer Vision and Machine Learning used in StreetScore?
Previous research in computer vision and machine learning has shown that a combination of human-labeled data and image features can be used to predict beauty1, interestingness2 or memorability3 of images. StreetScore makes novel contributions to this research in two ways. First, we extend the idea of predicting emotional response from images to predicting perceived safety. Second, and more importantly, we generate map visualizations by predicting the perceived safety of thousands of street view images from the same city. These visualizations are highly useful for urban science research. The key idea here is that we can aggregate the information obtained by an algorithm from images in a large collection to generate meaningful insights about the image collection.

As big visual data becomes common in social media, urban science, public health, entertainment and many other fields, computer vision algorithms that can synthesize information from troves of visual data will become important. We are excited to work on algorithms and applications for this fast developing field.
4) How is StreetScore useful for research in urban science?
First, StreetScore allows us to quantitatively explore dimensions of the urban environment that have been difficult to study. For instance, in the future we are planning to use StreetScore to study the dynamics of urban improvement and decay. We can also use StreetScore to study the spatial segregation of the urban environment (which is a form of urban inequality). Second, we can use StreetScore to explore the determinants of urban perception, including both, the local features that affect people's perception of a street view (such as the quality of buildings, streets and landscaping), the local context of a neighborhood (such as its proximity to a highway, railroad track or water), and the historical, economic and demographic forces that shape a city (such as architectural movements, temporal booms in construction, and the dynamics of ethnic and social segregation). Finally, StreetScore can empower research groups working on connecting urban perception with social and economic outcomes by providing high resolution data on urban perception.
5) How accurate is StreetScore in predicting the perceived safety of a street view?
As a binary classifier — designed to separate images that score less than 4.5 from those that score more than 5.5 — StreetScore is 84% accurate. On average we find that StreetScore explains 54% of the variance of original scores when predicting a value between 0 and 10. Please see the paper for a detailed evaluation.
6) This image looks quite safe to me. Why does StreetScore give it a low ranking?
StreetScore can predict a low ranking for a safe looking street view when evaluating images with visual elements it has not encountered in the training dataset, which consisted of only 3,000 images from Boston and New York City. These errors emerge in part because our training dataset is not comprehensive enough for StreetScore to learn all of the visual variation found in urban environments. Also, the scores presented were estimated using the images available on Google Street View at the time of scoring (spring of 2014). Google Street View images change over time and discrepancies between images shown and their scores may involve images that have been updated.

You can help us improve StreetScore by contributing training data. Place Pulse 2.0 is now collecting data on 100,000 street views from 56 different cities on seven different questions. You can contribute training data here.
7) Is the perception of safety connected to crimes?
The connection between crime and urban perception is a long standing question, which can be traced back to Wilson and Kelling’s broken windows theory4. As you can imagine, plenty of research and debate has surrounded this question and arguments and evidence have been collected in favor5, and against6 Wilson and Kelling's theory. In a previous paper some of us found a significant correlation between the perception of safety and the number of homicides that took place in a NYC zip code, after controlling for the income, population, area and age of each zip code. By improving the ability to collect data on urban perception we hope to contribute to this ongoing debate7.

Finally, we would like to remind visitors to this site that crime is a complex issue involving environmental, social8,9, and psychological factors. Visitors of this site should interpret these maps in the rich context provided by this multitude of factors.



1. Ritendra Datta, Dhiraj Joshi, Jia Li, and James Z. Wang. "Studying aesthetics in photographic images using a computational approach." European Conference in Computer Vision(2006)
2. Sagnik Dhar, Vicente Ordonez, and Tamara L. Berg. "High level describable attributes for predicting aesthetics and interestingness." Computer Vision and Pattern Recognition (2011)
3. Phillip Isola, Jianxiong Xiao, Antonio Torralba, and Aude Oliva. "What makes an image memorable?." Computer Vision and Pattern Recognition (2011).
4. James Q. Wilson and George L. Kelling. "Broken windows." Atlantic Monthly, 249.3 (1982): 29-38.
5. George L. Kelling. "Fixing broken windows: Restoring order and reducing crime in our communities." Simon and Schuster (1996).
6. Bernard E. Harcourt. "Illusion of order: The false promise of broken windows policing." Harvard University Press (2001).
7. For a cool experimental study showing evidence in favor of the broken windows theory check: Kees Keizer, Siegwart Lindenberg, and Linda Steg. "The spreading of disorder." Science 322.5908 (2008): 1681-1685.
8. Robert J. Sampson, Stephen W. Raudenbush, and Felton Earls."Neighborhoods and violent crime: A multilevel study of collective efficacy." Science 277.5328 (1997): 918-924.
9. Robert J. Sampson, Jeffrey D. Morenoff, and Thomas Gannon-Rowley. "Assessing neighborhood effects: Social processes and new directions in research." Annual Review of Sociology (2002): 443-478.