Multi-Plane Projection for Extending Perspective Image Object Detection Models to 360° Images

Abstract

Since 360° cameras are still in their diffusion phase, there are no large annotated datasets or models trained on them as there are for perspective cameras. Creating new 360°-specific datasets and training recognition models for each domain and tasks have a significant barrier for many users aiming at practical applications. Therefore, we propose a novel technique to effectively adapt the existing models to 360° images. The 360° images are projected to multiple planes and adapted to the existing model, and the detected results are unified in a spherical coordinate system. In experiments, we evaluated our method on an object detection task and compared it to baselines, which showed an improvement in recognition accuracy of up to 6.7%.