Training of Camera Encoder

[MnKnight1]

The results of your work look very good! But I have a question. In my opinion, the function of Camera Encoder is to perceive the perspective of an object, while the function of consistency loss is to make Camera Encoder more effective. But from the paper, it seems that Camera Encoder has to be retrained every time an object is generated, rather than being applied to any object after the training on a data set. I think a Camera Encoder should be suitable for any object, instead of retraining every time.