loss_fcos_ctr is always around 0.6

[Asahi678]

These days I trained the centermask(lite_mv2 and vov99) model on coco and my own dataset. However the training can not make loss_fcos_ctr go down. The loss_fcos_ctr is just around 0.6, the total loss is even get about 0.7. The two training experiment get the same result and I got my data checked again to make sure it correct. Now, I don't know how to solve this problem. Is there anyone else get the same problem like this?

[choodly]

I also encountered the same problem with my own dataset, how do you solve it?

[Asahi678]

I also encountered the same problem with my own dataset, how do you solve it?

Actually, I haven't solved this problem until now. But I find pred centerness is only used in calculating centerness loss. It is not involved in NMS. Box reg loss is calculated with target_cterness. I think the source of this problem is probably here, but I need more experiment to verify this.

[caganselim]

I have the same problem, even I have overfit to a single batch (~60 images) and got a high AP.

Maybe the authors can share the loss curves in tensorboard format so that we can analyze our loss curves?

[c-axel]

i have the same issue, can we just ignore this loss?

[ousinkou]

Same problem. I guess it's the bug of FCOS.

[GrassBro]

I implemented blendmask and centernet, both built on FCOS, on my own dataset, and I got the same problem. I am still new to FCOS, see anyone has a good suggestion here.

[zeeshanalipanhwar]

Answer: tianzhi0549/FCOS#15 (comment).