Zero loss_box_reg

Hello,

I am trying to train on my custom dataset (only one class, one bbox per image, no segmentation). My goal is to predict the correct class and the bbox after training. I followed the tutorial of detectron2 but I see a difference when I train: the loss_box_reg is always 0. Also, the prediction after training works

For example: total_loss: 0.850 loss_cls: 0.848 loss_box_reg: 0.000 loss_mask: 0.000 loss_rpn_cls: 0.002 loss_rpn_loc: 0.000

Do you have an idea why it is so?

[cybaj]

In my case, this results occurred when my annotations have it’s category_id property which starts from 1, not 0. Maybe you should check this out.

You can find the documentation about this dataset property here.

[KangolHsu]

I remove empty “segmentation” value in coco format json file，then it goes right now