An interesting experiment on how the thres_map affect results

[leidahhh]

**作者您好，我对自适应阈值一块比较感兴趣。在实验中发现一个比较有意思的现象，就是我把阈值损失以及binary的损失都删除了，只保留了模型预测的损失，结果模型的整体性能有了很大的提升。我想跟您讨论一下这个现象，以及您当初设计这个模块的想法，期待您的回复
Hello, the author. I'm interested in adaptive threshold. An interesting phenomenon was found in the experiment, that is, I deleted the threshold loss and binary loss, and only retained the loss predicted by the model. As a result, the overall performance of the model has been greatly improved. I want to discuss this phenomenon with you

删除前的结果：
2022-09-14 07:58:42,295 DBNet.pytorch INFO: [287/1200], train_loss: 0.4967, time: 133.9059, lr: 0.0007836829637320193
2022-09-14 07:58:45,779 DBNet.pytorch INFO: FPS:30.785972625664495
2022-09-14 07:58:45,780 DBNet.pytorch INFO: test: recall: 0.458333, precision: 0.964912, f1: 0.621469
删除后的结果：
2022-09-28 09:09:11,810 DBNet.pytorch INFO: [287/1200], train_loss: 0.1195, time: 145.5552, lr: 0.0007836829637320193
2022-09-28 09:09:33,585 DBNet.pytorch INFO: FPS:34.50438997451759
2022-09-28 09:09:33,589 DBNet.pytorch INFO: test: recall: 0.762254, precision: 0.931540, f1: 0.838438