search

cv516Buaa / tph-yolov5

GNU General Public License v3.0
716 stars 170 forks source link

关于验证集与测试开发集上的推理结果相差较大的问题 #32

Open Xavier-Lin opened 2 years ago

Xavier-Lin commented 2 years ago

使用您所给的权重进行推理，在验证集（40.8）和测试开放集上的结果（32.2）相差很大，特别是在测试开发集上的推理结果（32.2）与论文结果有一些区别，是我自己哪里设置的不对吗，想请教一下您这个问题，感谢。 测试开发集的推理结果 验证集的推理结果

cv516Buaa commented 2 years ago

It is normal for the results of test-dev set to be lower than that of val set. The accuracy of tph-yolov5 on the visdrone dataset is val > test-challenge > test-dev. The results of test-dev set shows in paper are obtained by training the tph-yolov5x model on train and val set, so the results obtained are slightly better than the given weights v5l-xs-2 and reaches 34.90.

Xavier-Lin commented 2 years ago

Thank you for your explanation of the accuracy of different data sets