why I can not training with 2d dataset 'PennAction' and 'PoseTrack'

[ZhangHedongTimes]

when i am training with 'PennAction' or 'PoseTrack' dataset , i always get a result" MPJPE error is 189.0999972820282, higher than 80. Exiting!..." 。 Is there some bug with the dataset?

[thinkingIsMagic]

我也是，同学你解决了吗

[ZhangHedongTimes]

解决了，里面有两个bug，一个是拿数据的时候需要深拷贝（添加一个copy），否则对2D keypoints 进行归一化时会使数据集也跟着边。还有一个就是关键点对应错误的问题，你可以把顶点的label绘制出来看看。

[SatoshiYamazaki]

@ZhangHedongTimes Thank you for your information! I already have found a bug on https://github.com/mkocabas/VIBE/blob/master/lib/dataset/dataset_2d.py#L68. We can fix the repeated normalization by adding np.copy().

 kp_2d = np.copy(self.db['joints2D'][start_index:end_index+1])

Do you find another bug related to this issue?

[ZhangHedongTimes]

The 'convert keypoints' is incorrect for 'pennaction'， on https://github.com/mkocabas/VIBE/blob/master/lib/dataset/dataset_2d.py#L70

[halcyon370]

It seems there is left-right inversion problem for 'pennaction'.

[zhanghan174]

The 'convert keypoints' is incorrect for 'pennaction'， on https://github.com/mkocabas/VIBE/blob/master/lib/dataset/dataset_2d.py#L70

hi~ I have the same error. May I ask how did you solve the problem of 'convert keypoints' about 'pennaction' dataset?

[hanabi7]

@zhanghan174 Did you solve the issue for pennaction dataset?

[hanabi7]

if self.dataset_name == 'pennaction': kp_2d = convert_kps(kp_2d, src='common', dst='spin') I add the following code in dataset_2d.py, but it don't seem to work

[winshot-thu]

@hanabi7 Have you fixed this bug? I have tried with the similar code but it do not seem to work too.