Training from scratch = no detections

yijingru / BBAVectors-Oriented-Object-Detection

[WACV2021] Oriented Object Detection in Aerial Images with Box Boundary-Aware Vectors

MIT License

466 stars 87 forks source link

Training from scratch = no detections #90

Closed candyotter closed 3 years ago

candyotter commented 3 years ago

Hello,

Everything works perfectly when using the pretrained weights downloaded from Dropbox. However, when training from scratch or resume training (on DOTA) and then running test or eval, the detections are always empty. They do not show on the image nor do any results appear in the merge_dota or results_dota txt files.

Do you know how I could solve this issue?

Thank you

kleinicke commented 3 years ago

You could lower the confidence threshhold (a commandline argument --conf_thresh). It might be too high, if the network is more uncertain with it's predictions.

candyotter commented 3 years ago

Thank you for your reply, I tried this but it didn't change anything unfortunately.

yijingru commented 3 years ago

Did you use the 600x600 cropped images?