the inputs of the "refinement model" are different between train.py and test.py about magnet HOT 6 CLOSED

Hi, as mentioned in the paper, to speed up the training process, the refinement module was trained to refine the whole output from the backbone network. In inference or testing, we will use the trained refinement module to correct progressively the about from preceding stages.

from magnet.

Hi, Thank you for your reply. However, I don't understand why the refinement model is trained on the input source (the output of the backbone model) while it is tested on the different input sources (the output of the preceding stages of the refinement model). So, these inputs have different distributions. With regarding IOU performance, Can it work in this way? What is the purpose of training refinement model?
Can you explain them to me? Thank you!

from magnet.

I think you misunderstood some points. The testing is still performed on the output of the backbone. After the refinement step, we only refine a part of the output of the model. The output is then used for the next stage.

from magnet.

Thank you for your answer!
Of course, The testing is still performed on the output of the backbone. Sorry, my question makes you confuse. I will simplify my question as follows:

in train.py

Output= refinement(A,B)

in testing.py

Output= refinement(A,C)

• A is the output of the backbone.
• B is cropped from the output of the backbone.
• C is the output of the refinement model of the previous stage.
  So, B and C are totally different. Is my understanding true? Is it ok if B and C are different?
  Sorry for my limited understanding.
  Thank you!

from magnet.

C is not totally the output of the refinement model of the previous stage. The output from the backbone (O) in the first stage is kept through stages. At each stage, a small portion of the output O is replaced by the output of the refinement module (to correct errors).

from magnet.

Thank you for your answer!

from magnet.