Paste-and-Cut: Collective Image Localization and Classification for Real-Time Multi-Camera Object Detection

Abstract

In recent years, object detection has emerged as a crucial task in various real-world applications, including security surveillance, autonomous vehicles, and robotics. However, traditional object detection models face numerous challenges, such as inefficient image processing, inadequate resource utilization, and a failure to consider the different criticality of input images, making it difficult to apply these models for timely inferences in practical applications. To overcome these challenges, this paper proposes a novel object detection framework, called Paste-and-Cut, that utilizes two techniques, image merging (paste) and RoI patching (cut), to optimize resource utilization and improve object detection performance. Additionally, our approach incorporates a dynamic merge size and canvas size decision mechanism to adapt to varying object detection environments. Experimental results obtained from experiments conducted with the MOT dataset demonstrate the effectiveness of our approach in achieving real-time object detection with improved detection accuracy and without generating any deadline miss. As such, Paste-and-Cut provides a promising solution for efficient and accurate real-time object detection in multi-camera scenarios.