A Picture is Worth a Thousand Dollars?

CMU B School Researchers Use Bridges to Unpack How Photos Help Sell Property

It’s not hard to believe that when you advertise a property, a good picture will help you sell or rent it. What isn’t so clear is what factors make for a good picture and how much each of these factors contributes to the bottom line. Using PSC’s Bridges, researchers at the Tepper School of Business at Carnegie Mellon have created a computer program that identifies these elements and shows how to present a photo for the best sales impact. 

 

Why It’s Important
 
The “global sharing economy”—in which nonprofessionals offer paid services via the Web—could add as much as $335 billion to the world’s economy by 2025. Today, Airbnb, which allows people to rent rooms in their homes online, has a corporate dollar value 20% higher than Marriott and hosts 25% more guests per night than Hilton. But because its hosts are not hospitality professionals, buyers face more uncertainty over the quality of a given lodging than with a known hotel chain. It’s not hard to believe that photos of a room can help rent it—or that a good photo is better than a bad one. But what makes a “good photo?” Will professional photography help sell a room, or will a “slick” photo seem phony and scare people a way? PhD student Shunyuan Zhang, working with advisor Param Vir Singh and colleagues at the David A. Tepper School of Business at Carnegie Mellon University, turned to high performance computing via PSC’s Bridges supercomputer to dig into the details.
 
“What’s the impact of images on the demand for a property in a listing? A good picture will increase demand, but by how much? What are the features in images that we should be focusing on? … We’d like to see, without restricting any dimensions; we’ll let the data tell us what are the dimensions that matter.”—Shunyuan Zhang, Tepper School of Business
How PSC and XSEDE Helped
 
Working with Dokyun Lee, Param Vir Singh and Kannan Srinivasan of the Tepper School, Zhang drew from research on art and professional photography, consumer behavior and psychology to identify 12 quality factors that are important in real-estate photographs. Using PSC’s Bridges, she set out to investigate whether scoring well in these factors in 380,000 Airbnb pictures led to better income for renters in 13,000 listings in seven U.S. cities. The 12 factors fell into three major categories: composition, color, and the relationship of important elements in a photo to the background. In a paper she presented at the International Conference on Information Systems last year, Zhang reported that photos taken by professionals certified by Airbnb increased demand by about 7 percent, earning renters an average $4,141 extra a year for a given property. Over 50 percent of this effect, she found, was associated with superior scores in the 12 quality factors.
 
In the next step in her work, Zhang went beyond the quality factors that had been identified by human experts. Instead, she allowed a “deep learning” program she devised to analyze 22,000 images of four room types and exteriors picked at random from a home-renting website.
The program taught itself by trial and error which photos associated with better sales—and which factors made the difference. To do this, it used the power of Bridges’ graphics processing (GPU) nodes as well as its CPU nodes to analyze the massive pixel-by-pixel data.
 
Importantly, the program identified the same factors already considered important by human experts. But it identified many more in addition, as well as returning very fine-grained measurements of how these affect sales. In all, the program identified a set of image features that could affect demand for a property. Some, such as the angle of the shot or the color of the bedsheets, need to be considered before photography. But others, like level of illumination and color tone, can be adjusted after the photo is taken and might eventually be included in automated image processing software.
 
“Bridges really helps us because of its computational power. Now, we are dealing with images and need a complicated model … that really consumes GPU and CPU power and memory. On another system the full analysis would take two months; on Bridges we expect it to take about two weeks.”