Experiments with Neural Net Object Detection System YOLO on Small Training Datasets for Intelligent Robotics

Abstract

In this paper we’ve conducted multiple experiments with modern object detection system YOLO. Object detection systems are fundamental to many robotics tasks. Recognition algorithms involving object detection are often part of various intelligence systems for robots. Training object detection systems usually requires waste amounts of training data which can be expensive and time-consuming. In this paper we’ve conducted several experiments with YOLO on small training datasets investigating YOLO’s capacity to train on small number of examples. We measured accuracy metrics for object detector depending on the size of training dataset, compared training process of full and smaller versions of YOLO and their speed. Gathered information will be used for creating visual factographic intelligence system for robots. YOLO (You Only Look Once) is a special intelligent technology for computer vision techniques. Our results are useful for industry professionals and students from a broad range of disciplines related to robotics, intelligent technologies and other fields.