Why I don't need to train TSP network seperately .

[liquor233]

Hello everyone, I am a beginner in ocr. When reading this wonderful code, I encountered a question: why TPS network can be trained together with the whole network. I mean, you don't know whether the real fiducial points are detected by the network. I have some experiences in face recognize, the face detect and face landmark detect are always seperated from face recognition network. So can someone explain to me why we don't need to train TSP seperately ?