such prediction boxes arise within CornerNet-squeeze result

[DongDongXA]

it is clear that the bound boxes' size are wrongly identified for that the cornerpoints are wrongly grouped.
is there any ways such as nms configurations rather than modifying networks to avoid such kind of mistakenly predicted boxes?

[heilaw]

The network mistakenly groups corners from different objects. You can try lowering the ae_threshold in the configuration file. If the distance between the embeddings of two corners is above ae_threshold, that pair of corner is rejected.

[pingqi]

@heilaw ,when to lower the ae_threshold? training or testing? or both? when I am training, the ae_thresthold is 0.5, if I using the model to test, and set the ae_threshold is 0.3, can it work?

[float4189]

I also encountered the same problem, did you solve it?