many roads lead to… Rome?
object detection models might infer bounding boxes, or segmentation masks.
you could train one to predict the corner coordinates instead of just a bounding box. or you could have it infer an activation map with the corners being hot, but that could be tricky to group into detections if you have more than one card in view.
or train a model to infer the mask of the card, instead of its outline. that is a whole lot more robust.