An Efficient Projection Screen Detection from Presentation Videos using CustomYOLOv7 Object Detection Method

Abstract

Lecture videos are analyzed for the development of various applications that involves indexing, summarization, content extraction, search, and navigation. Lecture videos captured in classrooms and conference rooms has digital slides projected on to the screen on stage. Projection screen detection is a crucial task for the extraction of slide region in such presentation videos. In this paper, we present an interesting approach for detecting the location of slide region in video frames using the You Only Look Once (YOLO) object detection model. First we train the custom YOLOv7 model on a labelled dataset of frames from presentation videos showing projected slides and then the trained model is applied on new images that are not used in training to predict the location of projector screens. We collected and annotated over 2000 frames from various presentation videos and then various augmented techniques are applied to prepare a dataset of 5000 images. We evaluated this method on our custom dataset and the results are compared with other popular object detection methods. Our experiments demonstrated that our custom YOLOv7 model outperforms basic YOLOv7 and Retinanet with regards to accuracy and computational effectiveness. Our results suggest that custom YOLOv7 provides a promising solution for projector screen detection and has the potential to be applied in various practical applications