search

akhilpm / DroneDetectron2

Pytorch code for our CVPRw 2023 paper "Cascaded Zoom-in Detector for High Resolution Aerial Images"
MIT License
52 stars 7 forks source link

Problems during reimplementation #21

Closed starkLJ closed 1 year ago

starkLJ commented 1 year ago

First of all, thank you very much for your excellent work. I've been trying to replicate your results, but I've encountered some issues, and I hope you can help me with them. Firstly, the results I obtained during replication did not match the results mentioned in the paper (AP 33.02, AP50 57.87, AP75 33.09), and there is a significant difference. Below are the training and testing commands I used:

Training:

python train_net.py --num-gpus 1 --config-file configs/RCNN-FPN-CROP.yaml OUTPUT_DIR outputs_FPN_CROP_VisDrone

Testing:

python train_net.py --eval-only --num-gpus 1 --config-file configs/RCNN-FPN-CROP.yaml MODEL.WEIGHTS <your weight>.pth

I hope you can point out any mistakes I may have made. Thanks again.

akhilpm commented 1 year ago

Check the issue #16.

zuikeaideren commented 1 year ago

Can you share your R-50.pkl with me, thank you

akhilpm commented 1 year ago

It is not mine in any regard. I downloaded it from the source mentioned in Detectron2's official code.

zuikeaideren commented 1 year ago

image I downloaded R-50.pkl here, is it the correct pkl file?

akhilpm commented 1 year ago

Yes

zuikeaideren commented 1 year ago

Okay, then the difference in experimental results should be a configuration problem, thank you.