Low-Power Image Recognition Challenge

Mobile systems (such as smartphones and drones) become increasingly popular and they are equipped with cameras. It is desirable to provide computer vision technologies so that these mobile systems can detect and recognize the... [ view full abstract ]

Mobile systems (such as smartphones and drones) become increasingly popular and they are equipped with cameras. It is desirable to provide computer vision technologies so that these mobile systems can detect and recognize the objects captured by the cameras. Even though many studies have been published on the topic of low-power image processing and computer vision, there has been no system-level comparison of different solutions. In order to evaluate the state-of-the-art low-power image recognition, two competitions (called Low-Power Image Recognition Challenge, LPIRC) were held in 2015 and 2016. In LPIRC, each team brought a complete system that can retrieve images from a referee system through a network. Each image contains one or several objects in 200 categories (such as humans, tables, vehicles). The system has to identify objects by the categories and mark the locations of the objects using bounding boxes. Meanwhile, the power consumption of the entire system is measured. The score of each team is calculated as the ratio of the accuracy and the energy consumption. This poster will present the findings in the first two years.