train w/o pre-trained backbone

[StarsMyDestination]

Hi, thank you for this excellent work first!
Have you tried to train the model from scratch? I observed some strange things:

1. the mAP is much lower than reported (in fact mAP<30);
2. the init proposals not learned well, and I visualized it to find that the init boxes barely changed even the training process was finished.

Any ideas? thx a lot ^_^

[StarsMyDestination]

Actually, I printed the 'init_proposal_boxes.weight' of your provided model r50_100pro_3x_model.pth, and the boxes are not distributed as the paper says:

above image is snipped from paper.

the actual proposal boxes in the provided pth file, and it almost the same, obviously not randomly distributed on the image to cover the whole image area.

Is there any details not illustrated in the paper or I misunderstood something?
If not, how to explain this?

[PeizeSun]

Hi~
For training the model from scratch, we didn't try it. I guess it may need more training epochs.
For the init boxes, please see this issue: link.

[StarsMyDestination]

@PeizeSun thx for your reply.
I will first try the random init, if I find something I will leave a comment here,