About training on your own dataset

[zzzzyq-18]

Now there is a dataset of its own, no categories, only single-target tracking. The code for training and testing was changed according to the way of got-10k, and the training was normal, but the test effect was very poor. When the author changes to his own dataset, what should he pay attention to in training and testing?
train:

test:

（The metric calculations tested are borrowed from other models）