The Color of Tourism: Another Lens of Venice
As we have demonstrated in previous sections, urban form structural elements (such as landmarks) and the hospitality industry (such as Airbnbs) can help us understand and regulate tourism through a spatial and typological approach.
We now share some preliminary ideas on the use of color psychology theory in the context of the Venetian streetscape. While we acknowledge that the relationship between color and perception may be influenced by age, gender, and culture, we are interested in exploring the chromatic experience of tourists who navigate the city starting at Venezia Santa Lucia, the most common entry point to the historic city, and meander their way to Piazza San Marco. To do so, we’ve identified four common routes frequently trafficked streets, three of which are by land and one of which is by Vaporetto along the Grand Canal.
In future research, we hope to shed additional light on the visual pollution of the streets (Portella, 2016). The consequences that facade colors, commercial signs, and outdoor street furniture can have on the visual quality of urban areas and on people’s quality of life, could inform architectural, planning, and psychological perspectives.
Methodology
The first step was collecting data using the Google Street View API. Using the boundaries of the historic city, we made HTTP requests to the Google Street View API to extract three-dimensional panoramas across the years 2010 to 2019. Each image was then saved with its corresponding latitude, longitude, pano_id (a unique identifier from the API itself), and the year the image was taken. The image below an example of a panorama image.
We then used the OpenCV library’s basic object detection function to identify common objects in each panorama. In particular, we were interested in detecting people, since foot traffic is highly dependent on the time of day, season, weather, or other factors that we cannot control for in Google Street View. Below you can see the Street View image with overlaid bounding boxes and their corresponding object labels.
The third step was to implement sky detection across image dataset. In a similar vein, the sky in our pano images was also dependent on the time of day, season, and weather. To account for this variability, we utilized a robust sky detection algorithm, originally designed for autonomous robot navigation, to identify the pixels of each image that were most likely part of the sky.
After detecting both people and an estimated sky region in each image, we then removed those pixels from our area of analysis. The next image shows the Street View image with people objects and sky covered in black (0, 0, 0) pixels. In our analysis, we turned these pixels transparent (0, 0, 0, 0) in the RGBA color space to disregard those pixels in future steps.
Color Distribution
For each image, we generated a 3D scatterplot of all of the R, G, B values that appeared in the panorama. This helps us understand the general distribution of colors within an image. However, since every colorspace is continuous, our next step was to discretize these values by extracting general color “families” from the image.
Dominant Color
Using the RGB color distributions, we ran a k-means clustering algorithm to identify general color “groups” that were most prevalent in the image, then selected the highest-weighted cluster centroid as the image’s dominant color. The first image below shows this analysis with 5 clusters. The second has 30 clusters and accordingly shows a greater range of colors.
Colorfulness Index
Finally, we wanted to capture the “colorfulness” of each pano image, where a fully grey-scale image would be considered the lowest extreme of 0. This example image received a score of 173.27.
Results
Using the four routes (shown below), we then scaled this process up to analyze approximately 400 three-dimensional panoramas along these common routes of Venice. This interactive map can also be accessed here.
Discussion
While our ideal conclusion would have been to predict the level of “touristification” of different streets using these color metrics, we acknowledge that there are many factors that may influence one’s experience with color in the historic city. Primarily, many of the images varied based on the time of day, the weather, or even the foot traffic along the street at which each image was taken.
For example, note the effect of the time of day and weather of two images taken near the train station, Venezia Santa Lucia.
There are also technical limitations to this method of analysis. Basic 2D projections of panoramas can bias which objects, and thus, color, dominate the image based on its specifications (e.g. pitch, heading), whereas more sophisticated image stitching methods could remediate such effects. We also utilized simple object detection models to detect and subsequently remove people from each image. With more fine-tuned models, panorama data, and image manipulation, future iterations of this analysis may better pinpoint the true colors of the built environment of Venice.