DOEYK: Detecting Objects with Enhanced YOLOv3 and Knowledge Graph

Current state-of-the-art object detection algorithms are almost exclusively based on Deep Convolutional Neural Network (DCNN). These algorithms all require a large number of labeled examples for each of the object categories they can recognize. These algorithms will fail for novel objects that only very few or even no prior examples are available. These algorithms are also far less accurate when classifying partially occluded objects. In a life sciences laboratory, objects and equipment are often overlaying each other and being partially occluded, novel or custom-made equipment exists everywhere. To successfully detect and identify the laboratory equipment, IAI along with our collaborators, propose to design and implement an object detection framework. By incorporating an automatic knowledge graph into a modified YOLOv3 object detection framework, the proposed approach is capable of detecting novel and partially occluded objects and improving detection accuracy for all object categories. In Phase I, we successfully demonstrated the feasibility of the DOEYK object detection framework that leverages human knowledge distilled in knowledge graphs to help improve the object detection performance for laboratory equipment. In Phase II, we plan to build a full-fledged DOEYK prototype software that addresses DTRA’s needs.