Object Detection with YOLOv5 in Indoor Equirectangular Panoramas

Abstract

With growing interest in indoor environment digitization and virtual tour creation, the acquisition of spherical images is gaining importance. Today, these images can come from a wide variety of devices, from dedicated industrial scanners through semi-professional sports cameras to smartphones equipped with the proper software. They also constitute a crucial input for smart indoor systems, such as automatic vacuum cleaners or surveillance cameras, and are essential for the construction of a building's digital twin. One of the most popular data representations used for spherical pictures is an equirectangular panorama (ERP). Although popular, processing it is demanding - ERP is heavily distorted, typically of high resolution, and made with discontinued lighting conditions. Transferring it to the cloud is challenging and the in-place analysis on the edge device is complicated. Nevertheless, ERPs store massive amounts of data, especially in the context of the number of objects captured in a single image. Many techniques dedicated to aberration-resilient object detection were presented. However, they all required proper preprocessing of the data or changed the detection approach so much that it made them unfit for any other task. In this paper, we present research on the processing of an indoor equirectangular panorama (I-ERP) using a family of object detectors called YOLOv5 (You Only Look Once, fifth generation). We discuss the process of adjusting an already available dataset to a newer annotation representation and test the usefulness of these detectors on such input without any changes to the detector models themselves. The results achieved indicate that the newer generation of YOLO detectors can be used with satisfactory results on the panoramic input and that the performance gap in I-ERP processing between dedicated and out-of-the-shelf solutions is closing.