The StreetScore Dataset
The StreetScore dataset contains scores of perceived safety for 551952 locations from Boston and New York. We calculated these scores by using machine learning and computer vision algorithms trained on less than 3,000 images from Boston and New York that were manually scored through a crowdsourced visual survey ( Place Pulse 1.0.). It is important that users of this dataset interpret it in the narrow context of the data collection method and its limitations.
Limitations
This data should be interpreted in the narrow context defined by the data collection method used for training the algorithm. This is data on images captured from a vehicle with a 900 field of view, and hence, represents the way cities look from a car. Moreover, most of the images were captured early in the morning, and hence, show images of cities with empty sidewalks, little traffic, and many of the shops closed. Also, the images of each city were chosen along a uniform grid placed within a manually-defined boundary. The boundary was chosen to include a large portion of a city and its surrounding metropolitan area, but keeping in mind that more area would demand more computation. Also, the scores presented were estimated using the images available at the time of scoring. Google street view images change over time and discrepancies between images shown and their scores may involve images that have been updated. The scores we present in StreetScore are not calculated dynamically, but statically at a given time period (spring of 2014).
Description
The dataset is made available in the form of a single .zip archive which contains comma-separated value (CSV) files for each city. The CSV files have three fields — latitude, longitude and q-score. You can locate the image evaluated by StreetScore by querying the Google Street View Image API using the latitude and longitude. Q-score is the perceived safety value for an image as predicted by the StreetScore algorithm (see paper for details). Please note that q-scores have been converted to deciles in the maps that appear on the StreetScore website to ease interpretation.

Please cite the following paper if you use this dataset for your research —
StreetScore - Predicting the Perceived Safety of One Million Streetscapes
Nikhil Naik, Jade Philipoom, Ramesh Raskar and César A. Hidalgo. CVPR Workshop on Web-scale Vision and Social Media (2014)

⇣ Download StreetScore Data