YOLO-based Object Detection in Panoramic Images of Smart Buildings
Abstract: Collecting and analyzing spherical images is one of the crucial fields supporting the digitization of indoor environments, like houses, apartments, offices, or factories, and creating virtual tours of smart buildings. Such images also constitute an important feed for smart indoor devices, like autonomous vacuum cleaners, intelligent production lines, or assistive droids. However, smart devices must correctly identify internal objects on high-resolution spherical images, usually heavily distorted and made with variable lighting conditions. Moreover, the recognition task must be relatively accurate and take place onboard the device, as data transmission to larger analysis centers or workstations does not allow for real-time operation. In this paper, we compare two object detectors from the YOLO family (YOLOv5 and YOLOv8) on the publicly available dataset adjusted to a newer annotation representation that allows for adapting YOLO detectors to spherical images. We verify the effectiveness of these two families of detectors for varying sizes of object detection models, image sizes, and training batch sizes. Our research proves that YOLO models can successfully detect most indoor objects, and their effectiveness is comparable to dedicated detectors having much higher computational complexity.
Loading