Commercial Vacancies

Inferring storefront vacancy using mobile sensing images and computer vision approaches

Storefront vacancy has been a widespread and worldwide phenomenon, raising concerns about the changing characteristic of the retail landscape, loss of community vitality, and hollowing out of cities. Although the causes leading to this phenomenon have been extensively debated, little granular data are available to evaluate the issue in a timely manner. Therefore, this study aims to develop a data-driven approach to capture the commercial structure of vacant storefronts on a store-by-store basis as well as to analyze their evolution patterns. First, streetlevel images were collected using mobile sensing in a low-cost, large-scale and efficient manner; then, a storefront vacancy estimation model was developed using computer vision techniques to infer the storefront location, operation status, business category, and vacancy rates. Three volunteers spent five days collecting street-level images from an urban area of 964 km2 in the case city of Xining, China. As a result, 93,069 stores were identified in the city in March 2022, of which 25,488 were vacant. Moreover, the storefront vacancy rate increased significantly after the epidemic, from 21.8% in 2018 to 30.0% in 2022. Stores in shopping, catering, and life services had the maximum vacancies. The factors that had the greatest impact on storefront vacancy were, in order of importance, far from commercial zonings, low population density, and far from the urban center. However, these factors influenced the vacancy in diverse and complex ways, and in the future, urban planning strategies to address vacancy issues should be well considered and differentiated. 
Li and Long 2024 CEUS_StorefrontVacancy.
Adobe Acrobat Document 8.4 MB

This dataset functions as supplementary material for the paper entitled 'Inferring Storefront Vacancy Using Mobile Sensing Images and Computer Vision Approaches,' which has been published in the Journal Computers, Environment, and Urban Systems. The dataset comprises the pre-trained Faster RCNN model, meticulously crafted for the recognition of vacant shops (located in the modal_data folder), along with the corresponding training data formatted in VOC within the VOCdevkit folder. Additionally, the GIS results of identified stores and aggregated outcomes at the street level are stored in Results_Xining.rar. For a comprehensive understanding of usage guidelines, please refer to the detailed operational instructions outlined in the README.