The error about datasets about skeletonmerger

[eliphatfs]

AFAIK the dataset is delivered with train/test splits? What do you mean by divide?

[makangzhe]

数据集我手动把ShapeNetCore.v2 数据集分为/train /val 两个文件夹下，然后可以运行了。我有一些疑问，如果我要检测自己数据集里面的关键点，我需要怎么构建数据集？因为我对这个ShapeNetCore的数据集不是很熟悉，不知道拿来训练的.h5包含什么信息。 如果我的数据集里面有深度图,物体对应mask ,还有模型.ply文件，可以用这个算法来生成关键点嘛？希望得到您的回复，万分感谢！

[eliphatfs]

Inside .h5 is a collection of point clouds and their class labels, as documented.

Current Merger is for per-object keypoint detection, since it is designed to give keypoints with correspondence in category level. It is not suitable for a localization task. If you have localized objects you can feed sampled point clouds into Merger. You need to run some sampling (invoking open3d for example) on the .ply model files to obtain point clouds.

Though, perhaps we can adapt Merger for keypoint detection in-the-wild with some modifications. If you have any idea about it you are welcome to talk about them here. (I suggest opening a dedicated issue for that.)

[eliphatfs]

Closing due to lack of activity in a month's period. Start another issue or reply to reopen if necessary.