Zero-Shot Object Detection Using YOLO

摘要

Zero-shot object detection is an emerging approach that aims to detect unseen object categories by leveraging the knowledge from known categories. In this study, we propose a novel zero-shot object detection framework based on the popular YOLO (You Only Look Once) architecture. Our approach addresses the challenge of detecting objects that have not been seen during training by learning a mapping between the visual features of unseen categories and their textual descriptions. We introduce a semantic embedding space that captures the relationship between visual and textual representations and use it to generate visual features for the unlabeled object instances. By aligning these visual features with the textual descriptions of unseen categories, we enable the detection of previously unseen object classes. Our experimental results demonstrate the effectiveness of our approach, achieving competitive performance for both unseen and seen object categories on kaggle datasets. Moreover, our framework's incremental learning capability allows for the continuous incorporation of new unseen object categories without the need for retraining the entire system. The proposed zero-shot object detection method opens up new opportunities for various applications, such as autonomous vehicles, surveillance systems, and robotics, where the ability to recognize novel objects is crucial with the high level of Accuracy-98.2%, Precision-97.9<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup>, Recall-97.8<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup>, F1 score-96.8<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> using You Only Look Once (YOLO) in machine learning algorithm.