the classes num in dataset doesn't correspond to the parameter (bbox_head) "num_classes" in config file. Could you explain it :)

LiWentomng / OrientedRepPoints

The code for “Oriented RepPoints for Aerial Object Detection (CVPR 2022)”

268 stars 43 forks source link

the classes num in dataset doesn't correspond to the parameter (bbox_head) "num_classes" in config file. Could you explain it :) #25

Open Xiloy opened 1 year ago

Xiloy commented 1 year ago

for example, i find you use "wordname_15" dataset, the num classes of this dataset is 15, but in your config file, the parameter “num_classes" is 16. The values of these two numbers always seem to differ by 1...

LiWentomng commented 1 year ago

In the previous vesion of mmdetection (v1.x), the background class is included in the config file, i.e, "num_classes=15+1".

Xiloy commented 1 year ago

thanks! But I still have a problem that I haven't figured out yet. When i train the model, it prints infromation as Screenshot from 2022-10-26 18-55-56 I want to confirm (1) what is the difference between loss_rbox and lossspatial (2)Why should we need to set loss#_init and loss#_refine or what's their use thank you again!

LiWentomng commented 1 year ago

@Chenanxinyi (1) Loss_rbox is designed for points learning as the main term, loss_spatial is designed for the outlier point as the penalty item. (2) This is coase-to-fine framwork, the refine stage will further be better based on the init stage.