Description
Hi, thanks for your great work!
I have a question. You showed results on HPatches that compare the evaluation from different models in the table. The first row shows the evaluation from pre-trained model which is from SuperPointPretrainedNetwork, and the third row is from this GitHub.
I notice that repeatability(in Detector metric) from this GitHub is better than that from pre-trained model, but matching score(in Descriptor metric) is lower distinctly. The result above mean that model from this GitHub can detect feature keypoint better, but match worse because descriptors have a poor descriptive ability. I observe that you use different batch size, learning rate, lamda-loss(desc loss/detect loss) from pre-trained model in the training details. I tried to adjust same parameters mentioned above with pre-trained model, but match performed even worse.
So, I wonder why the result mentioned above occurred, and how to make the training reach agreement with the pre-trained model.
Looking forward to receiving your reply!