第一个训练指令已经是使用ESOD了吗，评估结果有gap

[DongRos]

如题，训练指令为DATASET=visdrone MODEL=yolov5m GPUS=2 BATCH_SIZE=8 IMAGE_SIZE=1536 EPOCHS=50 bash ./scripts/train.sh
评估指令为python test.py --data data/visdrone.yaml --weights runs/train/exp2/weights/best.pt --img-size 1536 --device 0,1
只评估了visdrone数据集，评估输出结果如下图

显示map只有34.5，与论文的36.0有点差距，是我哪里设置的问题吗，还麻烦作者帮忙看看，感谢！

[DongRos]

使用作者提供的预训练模型在验证集评估 map33.1
​​

在测试集评估 map25.3

是什么原因导致的与论文数据的差距？

​​

[tso2381637]

Hi, I got the same result as yours. Have you figure out what's going on?

[yuhua666]

I also got the same result as yours.

[judycpChen]

Maybe you can try passing in some other parameters when running test.py, such as --use-gt. When I run the original inference command given by the author directly, the accuracy value on the visdrone validation set is 34.6, but when I pass in the --use-gt parameter, the accuracy value is 35.3.

[andy690]

Maybe you can try passing in some other parameters when running test.py, such as --use-gt. When I run the original inference command given by the author directly, the accuracy value on the visdrone validation set is 34.6, but when I pass in the --use-gt parameter, the accuracy value is 35.3.

这样会造成测试集的数据泄露，肯定是不能拿来当指标的

[andy690]

作者的实验设置中，测试集数量显示为3190，而我们用的测试集数量都为1610，为什么会有这样的差异？