Real-time inference of vision system in robotics has recently acquired significant prominence in self-driving cars, drones, agriculture, food grading, warehousing and surveillance systems. These systems require fast and accurate interpretations of the vision data as delays and inconsistencies can have fatal consequences.
This repository is accompanying the Active inference for embedded vision systems paper that looks at the GoogLeNet architecture for real-time object classification, producing accuracy of 75% at an inference speed of 189 fps for use on resource-limited hardware systems, such as the Nvidia Jetson platforms’, along with different forms of data acquisition for extending datasets.