Yu Sun

可以简单通过[这样的](https://github.com/Arthur151/ROMP/blob/8606d695e104f550a83bd74da3f3927c8e01faaf/simple_romp/romp/post_parser.py#L90)实现转换。 是的，完整评测、可视化代码会陆续放出来的，赶dataset开源deadline没来得及搞全。

Sorry, this rep. is no longer maintained. To get these data, please refer to https://github.com/Arthur151/ROMP/blob/master/docs/dataset.md

Thanks for your interest! Yes, of course! But they won't be release in this repository. Limited by the licence agreement of some datasets and quoted codes, we may not provide...

Thanks for your interest! Yes, we save the SMPL parameter in the h5 file. which compose of 3-dim camera parameter, 72-dim SMPL 24 joint rotation vector, 10-dim SMPL shape parameter....

Actually, the root_rotation we provided is already considering the camera pose. So you don't need have to multiple the extrinsic matrix.

Please don't multiply any camera matrix, such as projection_matrix you mul. The rendered 3D mesh in the right is just captured by the camera placed in a wrong place. Please...

Thanks for your interest! Yes, of course! The training code will be released. But they won't be release in this repository. Limited by the licence agreement of some datasets and...

对于前两个问题，因为ＤＳＤ这篇文章的主要competitor就是ＨＭＲ－ｖｉｄｅｏ，所以采用和它一样的ＧＴ用于评测，主要用于性能对比。所有的设置都是为了与ＨＭＲ－ｖｉｄｅｏ的设置保持一致的。 关于h36m数据集，因为每一帧都是按照每五帧采一帧下采样出来的，每一帧有３个ＳＭＰＬ参数是３个真正临近帧的ＳＭＰＬ，随机采样是为了避免过拟合，因为真正临近帧动作基本一致。 对于３ＤＰＷ数据集的处理还是建议看ＲＯＭＰ的[相关部分代码](https://github.com/Arthur151/ROMP/blob/7400b34d1d69112625a2165c78b5c754943651c0/src/lib/dataset/pw3d.py#L113)。

我这里2D pose对的挺好的呀，没什么问题。  我好像是丢弃了每个视频前１５帧，应该是整个采样平移了５为倍数的帧数，你可以试一下。具体的太久远，忘记了。。。

恩恩，我中间迭代了几版，都忘记怎么搞得了。方便把数据处理代码（图片抽取）开源给大家么？