Creating data.pyd and detection.npz for our own dataset, so we could train the 4D-humans algorithm on our own dataset.

[ChennyDeng]

Hi,

Thanks for releasing the code. I would like to train your algorithm on our own dataset. I wonder how should we sort out our dataset in the correct format? In the dataset you provided, for each image, you created a data.pyd file and a detection.npz file. I wonder what are these two files and how do you create them? Thank you for your time and effort in helping us.

[shubham-goel]

Hi @ChennyDeng, thanks for asking!

This will become clear once we release the preprocessing code. Till we make that public, hope this explanation helps:

The .pyd file contains almost all the info needed for training (as seen here and here). This includes bbox, keypoints, pseudo-gt 3D fits, etc.

The .npz file contains the output of the detector, including masks (see here). We bundled it with the tars during dev but its masks are not needed for final training and could be skipped.

[danielajisafe]

In the meantime, @shubham-goel please is it possible to read the data.pyd files without using the webdataset library. Thanks for your help.

[HeroBaoYX]

你好

感谢您发布代码。我想在我们自己的数据集上训练您的算法。我想知道我们应该如何以正确的格式整理我们的数据集？在您提供的数据集中，您为每个图像创建了一个 data.pyd 文件和一个 detection.npz 文件。我想知道这两个文件是什么，您如何创建它们？感谢您花时间和精力帮助我们。

Have you solved this problem now? Looking forward to your reply

[luoxue-star]

@ChennyDeng Have you solved this problem? Looking forward to your reply.