Computer Vision Assisted Approaches to Detect Street Garbage from Citizen Generated Imagery

The basis of smart governance is to leverage state-of-the-art technologies to improve lives of citizens. With the rapid permeance of smart-phone technologies today, citizens are increasingly active now in collaborating with public officials for improved quality of life. However, for effective utility, public officials must be empowered with optimal tools that can best leverage citizen participation. In this paper, we present the design and details of computer vision techniques to automatically detect and localize street garbage from citizen generated imagery, and analyze the performance of multiple techniques. Our dataset is mined from (citizen-generated) images in the well-known 311 service deployed in San Francisco, which is actually a service citizens use to report civic issues. Using a dataset of 2, 500 images (containing 6, 474 objects) evenly distributed between those containing street garbage and those that do not, we design and compare convolutional neural network techniques to detect and localize sources of garbage in the images. Results from our evaluations show that our system can be a vital cog towards next generation smart governance systems geared towards cleaner and healthier neighborhoods. Since identifying, collecting and disposing of street garbage is a critical aspect of governance across the globe, we believe that our work in this paper is critical, timely and may have global impact.