One of the problems with equirectangular images is that a standard object dataset like COCO (Common Objects in Context) doesn’t play well with objects on the edge.
This paper talks about using COCO with YOLO.
Their solution is to use four sub-windows.
As this is a research paper, it seems unlikely we’ll be able to solve the problem with the THETA V to get full 360 object detection. We may be able to a reasonable 210 degrees of detection without too much work. I’m curious to learn how other people have progressed.
