FairMOT MOT17 训练和测试的提取的特征差距很大

想一下训练和测试会差距在哪里呢？

model.train(), model.eval()

- BN

FairMOT, JDE数据加载的预过滤

FairMOT/src/lib/datasets/dataset/jde_reid_conf.py

valid_idxs = (w > 4) & (h > 4) & (area / (area0 + 1e-16) > 0.1) & (ar < 10)
targets = targets[valid_idxs]
targets[:, 2:6] = xy[valid_idxs]
targets = targets[targets[:, 2] < width]
targets = targets[targets[:, 4] > 0]
targets = targets[targets[:, 3] < height]
targets = targets[targets[:, 5] > 0]

数据增强！

我其实跑了很多次，但是没有测出什么区别来

原来是有数据增强，而且我锁死了随机种子！

数据本身的差距太大