Finding the Perfect Cut: Selection of the Best Cutting Point in Equirectangular Panoramas for Object Detection

Sebastian Pokucinski, Katarzyna Filus, Dariusz Mrozek

Abstract: Analyzing equirectangular imagery becomes increasingly important due to its expanding utilization in numerous domains. With their wide field of view and unique format, they create novel opportunities and challenges for object detection algorithms. Training on rectilinear data, which is better accessible than panoramic data, results in low accuracy on equirectangular images. Therefore, it is important to propose new dataset preparation methodologies that allow to fully utilize the potential of the size-constrained, relatively unpopular and niche equirectangular datasets. In this paper, we proposed a novel method for the selection of the Best Cutting Point that can be used to create datasets aimed at facilitating the model’s training and testing. This approach is used to choose the meridian from which the sphere is unfolded during the transformation from a 3D spatial figure into a 2D fat image. Our target is to eliminate the disadvantages of the existing cutting methods: the time-consuming, labor-intensive, and subjective character of a manual technique and frequent distortions in the spatial placement of object features common for a random cutting method. It does that by automatically finding the best cutting point along the horizontal axis to preserve the integrity of objects relevant to the detection task, either keeping them intact or minimizing losses. To evaluate the method, we use an available dataset and also propose a new dataset called EquiB&B. The numerical results show that the datasets created with the proposed approach can facilitate the training of detection models for equirectangular data. They also show the potential of the method for building practical testing methodologies involving best- and worst-case scenarios to introduce more explainability and transparency to testing.