About detecting many extra boxes

[pengjizhi]

Hello author, thank you for your hard work!
Drawing on your ideas, I am developing my own single-stage object detection network, but there are a lot of extra boxes in the visualization results.
Since my network structure is deep, I guess it may be caused by the disappearance of the gradient or the explosion of the gradient, but why do these two situations lead to the appearance of a large number of abnormal boxes, can you give any suggestions?

In addition, I don't think this is the cause of nms. In fact, such results are already the results after nms.

[hangZhou6]

hello你好, 请问你解决了吗？

[gordenxu]

(1)in this case, you can lower the CLS_THRESH = 0.5 can be higer, such as 0.9, while the NMS_THRESH = 0.5 can be lower, such as 0.1