softnms used in the training phase?

[liangzimei]

in my person detection task, when i use softnms in testing, the ap drops a little.
in default, is(should) the softnms used in training.?
thank you....

[ds2268]

soft-nms is used completely independently of the model that has been learned as it only processes/filters detections based on their scores and IoU scores (overlapping measure). There is possibility of slight drop in performance, try setting higher detection threshold compared to the one used by standard NMS, basically play with NMS threshold and detection threshold as this is your only chance to improve the results with soft-nms.

I suggest reading the paper for further clarification.

[hawkjk]

I try to add softnms in training, it seems that the result drops a lot. It's strange.
Now I only use softnms in testing , it seems that the result will be a little better in occlusion case but a little worst in no occlusion case.