Object detection on panoramic images based on deep learning

Panoramic image can be widely used in many applications, such as virtual reality, visual surveillance and autonomous vehicle, because of its large field of view. However, the inherent distortion for panorama causes object detection to be a challenging task. This paper focuses on the multi-class objects detection in panoramic images using deep learning method. The proposed system uses three fisheye cameras to efficiently create panoramas and build a large dataset. A region based convolutional neutral network (R-CNN) is implemented to train and test on an indoor panoramic image dataset. Experiments show great improvement performance on ten categories of distorted indoor objects with a mean average precision of 68.7%.