search

ucas-vg / P2BNet

ECCV2022, Point-to-Box Network for Accurate Object Detection via Single Point Supervision
MIT License
63 stars 9 forks source link

P2BNet on DOTA #11

Closed lo202ng closed 1 year ago

lo202ng commented 1 year ago

I have tried to train on DOTA using the COCO config, but got a bad result, the minimum loss displayed when training was more than 0.1, and the pseudo box in visualization is terrible. Does this model not apply to remote sensing data sets? these are one pic of my results and train information: 00004

image

Darren-pfchen commented 1 year ago

The anchor setting and learning rate can be changed in DOTA to adapt to the object scale. We have no time to adjust it now and we will try it in following work. Thanks!

lo202ng commented 1 year ago

thanks for your replay, I'll try it~