MCG-NJU / MixFormer

[CVPR 2022 Oral & TPAMI 2024] MixFormer: End-to-End Tracking with Iterative Mixed Attention
https://arxiv.org/abs/2203.11082
MIT License
448 stars 75 forks source link

关于一些实验结果的疑惑 #61

Closed Old-detective closed 1 year ago

Old-detective commented 1 year ago

您的论文中的Table 2里边显示的Mixformer的速度是25FPS,GPU是1080Ti,请问一下您这个结果用了几个GPU?是在哪个数据集上跑的?还有就是我用的Mixformer在UAV123数据集上跑出来的结果是6.1FPS,两个2080Ti,请问这正常吗?十分感谢

fire717 commented 1 year ago

同样好奇,论文中只有精度对比,没有放速度对比,一般不都应该放一个精度-速度的坐标图吗

yutaocui commented 1 year ago

跟stark等方法测试fps的方式一样的,参考profile_mixformer.sh,用单张卡进行测试,而且最好不用有其他程序同时在跑。我们尝试了vit和convmae等,速度上会提升很大。 论文中我们也report了速度