How is the class-agnostic detection score carried to the 2nd stage?

[qianyizhang]

can you show me which part of the code implements this? (what you claimed in paper section 4-Detector design)

the way i see it is that: you replace traditional 2 stage detector's RPN with a strong proposal generator. and I don't see how it suddenly makes the network probabilistic.

[QZLearning]

Hi，have you figured this out ? I'm still confused about this

[JimKlingshirn]

I think the code you're looking for is in custom_roi_heads.py.

CenterNet2/projects/CenterNet2/centernet/modeling/roi_heads/custom_roi_heads.py

In particular, the function _forward_box() retrieves the proposal scores and multiplies by the class scores.

[bravezzzzzz]

I think the code you're looking for is in custom_roi_heads.py.

CenterNet2/projects/CenterNet2/centernet/modeling/roi_heads/custom_roi_heads.py

In particular, the function _forward_box() retrieves the proposal scores and multiplies by the class scores.

Hi! Does this mean the class-agnostic detection score is not carried to the 2nd stage at training?

[sh7jacobi]

Hello, I use tensorboard and I have observed that there are stage0 metrics. What is stage0?

[chuong98]

Hi! Does this mean the class-agnostic detection score is not carried to the 2nd stage at training? @xingyizhou Can you help to clarify this question ?

[GrantorShadow]

Hello, I use tensorboard and I have observed that there are stage0 metrics. What is stage0?

Stage 0,1,2 are the three cascaded heads in Cascade RCNN

[GrantorShadow]

I think the code you're looking for is in custom_roi_heads.py. CenterNet2/projects/CenterNet2/centernet/modeling/roi_heads/custom_roi_heads.py In particular, the function _forward_box() retrieves the proposal scores and multiplies by the class scores.

Hi! Does this mean the class-agnostic detection score is not carried to the 2nd stage at training?

https://github.com/xingyizhou/CenterNet2/search?q=MULT_PROPOSAL_SCORE

Multiply proposal score is set to true, I suppose that's how the score is carried forward in the training stage. @xingyizhou Please could you help clarify?

[sh7jacobi]

Hello, I use tensorboard and I have observed that there are stage0 metrics. What is stage0?

Stage 0,1,2 are the three cascaded heads in Cascade RCNN So, CenterNetv2 is combined with Cascade RCNN?

[GrantorShadow]

Hello, I use tensorboard and I have observed that there are stage0 metrics. What is stage0?

Stage 0,1,2 are the three cascaded heads in Cascade RCNN So, CenterNetv2 is combined with Cascade RCNN?

It uses Cascade RCNN heads, Please refer to the paper for further details

[zhaoxin111]

[jehovahxu]

I'm also confused about this. It seems like only multiplies in inference

[jshilong]

I'm also confused about this. It seems like only multiplies in inference

Seems the code only multiplies the objectness of rpn to the second stage in the inference phase， which is in contrast to the first impression of reading the paper.

[jehovahxu]

I'm also confused about this. It seems like only multiplies in inference

Seems the code only multiplies the objectness of rpn to the second stage in the inference phase， which is in contrast to the first impression of reading the paper.

It multiplies the probabilistic of first-stage to the score of second-stage and it is the same as in the paper. But it is only used in inference and it seems like not mentioned it used in training in this paper.

[nilskk]

So it only multiplies the objectness score with the classification score during inference, because I also got the impression reading the paper that the multiplication would happen during the training phase? @xingyizhou can you maybe help out?

[wusize]

So it only multiplies the objectness score with the classification score during inference, because I also got the impression reading the paper that the multiplication would happen during the training phase? @xingyizhou can you maybe help out?

To my understanding, the objectness score should be the exponent to the background score. Anyway it's not implemented in the code. :)