Region Proposal Exploration for Extending Perspective Detection to Panoramic Detection

Since panoramic images provide a wide field of view of the surrounding environment, they offer a rich source of information that is increasingly utilized in fields like robotics and autonomous systems. With the application of 360° cameras (i.e. panorama, omnidirectional cameras), robots are gaining 360°perception, enabling them to better understand and interact with their surroundings. However, the unique characteristics of panoramic images, such as geometric distortion and spatial discontinuity introduced by the panoramic projection, pose significant challenges for the application of conventional object detection methods, which are typically designed for perspective images. To address these challenges, current approaches often rely on repetitive overlapping detections to avoid object fragmentation caused by view cuts, but this leads to detection failures and inefficiencies. In this paper, we propose a novel method that explores region proposals within panoramic images to perform view projection. By reducing redundant detections, our approach improves detection efficiency and accuracy. We demonstrate through experiments that our method enhances object detection performance, providing a more effective solution for panoramic image object detection tasks in robotics and autonomous systems.