This might not work if the cards have 6 degrees of freedom. You can check out CornerNet and its variants for anchor free corner estimation. The original paper detects two corners, but extending to four should be possible.

Another option is to use yolo to detect a rough bbox, and then use classical cv to refine corner locations.

https://github.com/XinzeLee/RotateObjectDetection

There are yolo versions that allow oriented bbox detection. Above is just one example, there is a lot of repos on gihub.