SUMMARY It is amazing what can be achieved with Deep Learning these days. The improvement in face detection with YOLO has been incredible, and it has brought this idea back to me again. Apply Deep Learning to search / detect images that are upside down (or rotated), and thus be able to later select all of them (or those that one needs / prefers) to correct them. I have been seeing that there are many projects in this regard, and quite well documented. Perhaps some owned "sanitized" albums (with correctly oriented images) could be used as training data for the neural network, and later applied to the rest of the library. In my tests, this would also improve facial recognition (not detection). Those images that are already tagged and that are still rotated, greatly reduce the precision when used as training data for recognition.
Maik, This is a task for the Neural Network used to auto-tag images. I don't know if the current models are able to detect this upside-down photograph. Gilles