Why do I get top1 and top5 very low when I train with train_bone.yam, only 3% and 8%? But when training with train_joint.yam, the top1 and top5 are close to the paper. I am using the kinetics data set.

[LiJieqin]

[lshiwjx]

You should provide more details

[dutlzn]

i meet the same question.

[LiJieqin]

i meet the same question.

someone told me the reason is we modified the open/close state，but I didn't validate if this is true

[dutlzn]

@LiJieqin in train process.Have you ever encountered a situation where Top1 equals zero

[LiJieqin]

@LiJieqin in train process.Have you ever encountered a situation where Top1 equals zero

no，which dataset are you using?

[dutlzn]

@LiJieqin i use the dataset like the kinetics

[LiJieqin]

you can try the solution I said before

[wangzeyu135798]

@LiJieqin 您好，您之前说的open/close state是什么意思？我也遇到了bone数据准确率很低的问题。