Human Pose Recognition in Top-View Fisheye Images

While top-view fisheye images are common in surveillance settings, the development of techniques for their analysis is far behind that of perspective images. In this paper, we focus on the problem of human pose recognition, which is rarely studied in such images, with the emphasis on the application of existing pose recognition models developed for perspective images to fisheye images. Both models that belong to the top-down and bottom-up approaches are included in the experiments. We investigate the use of model finetuning and virtual-PTZ views and show that they are useful for applying the existing models to fisheye images.