search

Chasel-Tsui / mmdet-aitod

Official implementation of AI-TOD-v2 and NWD-RKA
Apache License 2.0
52 stars 12 forks source link

About Visdrone2019 performance #4

Open JWZhao-uestc opened 2 years ago

JWZhao-uestc commented 2 years ago

Hello, Thanks for your sharing firstly. I download the VisDrone2019 and trained 20 epoches, my config file was followed Faster-RCNN*, but the AP I got is lower than the paper reported. Note that I rewrited a dataset file according to aitodv2.py in "mmdet/datasets", whether the file had someting wrong? The result I got (left) and the reported (right)： AP: 20.9 AP: 23.2 AP0.5: 39.5 AP0.5: 41.5 AP0.75: 19.3 AP0.75: 22.6 APvt: 2.1 APvt: 3.7 APt: 7.4 APt: 10.6 APs: 16.4 APs: 19.5 APm: 32.1 APm: 32.5 Hope for your replying, thanks.

Chasel-Tsui commented 2 years ago

Hi, thanks for your comments. Could you please provide the complete config file of the Faster-RCNN*, so that I can help check it?

JWZhao-uestc commented 2 years ago

Hi, thanks for your comments. Could you please provide the complete config file of the Faster-RCNN*, so that I can help check it?

Sure, the files are as follows mmdet/datasets/visdrone2019.py visdrone2019.txt base/datasets/visdrone2019_detection.py visdrone2019_detection.txt visdrone2019_faster_r50_nwdrka_1x.py visdrone2019_faster_r50_nwdrka_1x.txt Thanks!

Chasel-Tsui commented 2 years ago

There are two differences (the img_scale, and training epochs). I set the img_scale to (1333,800) in the train_pipeline and the test_pipeline for VisDrone, the same as that of COCO. The training epoch is set to 12, lr=0.005 with batch_size 2, and the lr decays at the 8-th and 11-th epoch. Besides, you could test the baseline performance first to see whether it is similar to that in the paper.

JWZhao-uestc commented 2 years ago

There are two differences (the img_scale, and training epochs). I set the img_scale to (1333,800) in the train_pipeline and the test_pipeline for VisDrone, the same as that of COCO. The training epoch is set to 12, lr=0.005 with batch_size 2, and the lr decays at the 8-th and 11-th epoch. Besides, you could test the baseline performance first to see whether it is similar to that in the paper.

Ok, the aitodv2 performance I trained is similiar to the reported. I will check my visdrone2019 config file according to your advice. Thanks for that again.