第一个训练指令已经是使用ESOD了吗，评估结果有gap

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！

Reproduce the stated training command from scripts/train.sh and the evaluation command in test.py using data/visdrone.yaml and the reported weights path. Compare the resulting mAP with the attached 34.5 result and the paper's 36.0, then inspect the training and evaluation settings to identify the mismatch. Done means explaining the gap or confirming the settings needed to match the paper.

Written by the indexing model from the issue text.