train on own data, loss cannot converge

[eternal-mind]

Training on my own data, the model does not use mid branch. Initially, the point loss will decrease, but after a period of training, it will rise to around 2.x, and oscillate near that level, unable to decrease. What could be the reasons causing this issue?

vis result shows the prediction boxex deviates significantly from the gt

[wplove]

Can I ask you a question？If I want to train my data set, how do I need to change my label format to ensure that the data is successfully read？