centerness shared head

[AllenPeng0209]

Hi,

I want to know that is there any experiment results show that centerness branch shared the same parameter with classification branch achieve the best MAP?

Would it be better to separate the centerness branch?

And what it the effect and result behind?

Thanks !

[tianzhi0549]

@AllenPeng0209 It is discussed in Section 4.1.2 in our paper.

[AllenPeng0209]

Also,

would it be better to train 80 centerness scores for 80 class instead of 1 score for 80 class?
Thanks

[tianzhi0549]

@AllenPeng0209 Sorry, we have not tried this. You might try it by yourself.

[bearcatt]

@tianzhi0549 I've tried sharing the center-ness head with the reg head rather than the cls head. This simple modification leads to 0.5 mAP gain for FCOS-R101. You can try this.

[tianzhi0549]

@bearcatt Thank you very much! We will try it.

[tianzhi0549]

@bearcatt We have confirmed it can improve the performance. Do you mind we add this into our code with a proper acknowledgement?

[bearcatt]

@tianzhi0549 that's great! You can also refer to the code here.

[tianzhi0549]

@bearcatt Thank you:-). Let me do it.

[18804601171]

@bearcatt Why I share center-ness head in reg branch at mmdetection,it's not improved

[JunsukLee]

The value of CENTERNESS_ON_REG in the provided yaml file is set to True.
The performance of the paper shows excellent performance when CENTERNESS_ON_REG:False.
If the default value of all yaml is True, is the performance excellent? I have confused.