Adaptive global reasoning with multiple knowledge graphs for object detection / Tao Bo

• Main Author: Tao , Bo
• Format: Thesis
• Published: 2021
• Subjects: QA75 Electronic computers. Computer science
• Online Access: http://studentsrepo.um.edu.my/14434/1/Tao_Bo.pdf; http://studentsrepo.um.edu.my/14434/2/Tao_Bo.pdf; http://studentsrepo.um.edu.my/14434/
• Institution: Universiti Malaya

The dominant object detection system's mechanism is to propose some regions of interest and then classify them and locate these regions with bounding boxes. In other words, the current object detection system is modeled as classification on boxes in parallel without considering the relationship between the objects. Such strong semantic information should be used to help current object detection systems to get more accurate results. In contrast, human vision recognition system can recognize objects easily, even in very complex scenes (heavy occlusion, more categories, class ambiguities, etc.). The main reason is that humans have the knowledge (common sense) to help them recognize what they see. When humans cannot see the target object clearly, the visual reasoning process goes on: With the help of surrounding objects and environment or context, humans usually have the ability to deduce the object. Inspired by the human visual recognition mechanism, many works have been done to incorporate knowledge base to current object detection system to imitate the reasoning process. The dominant reasoning process is to propagate region features through a fixed external knowledge graph. The nodes in the graph represent region proposals, and edges represent connections or relationships of each pair of nodes. After the learning process through the knowledge graph, the region proposals