Pretrainig skinning net does not converge on customized dataset

[bluestyle97]

Hi, I'm training SCANimate on ClothSeq dataset provided by Neural-GIF(https://github.com/garvita-tiwari/neuralgif). However, I met some problems and got bad results.

Here are the logs of the first stage when training on a Cape sequence:

And here are the logs of the first stage when training on a ClothSeq sequence:

I notice that the pretraining of the skinning net converges very quickly on the Cape sequence, while the loss drops slowly on the ClothSeq sequence. I have checked the minimal_body file, the scans and SMPL parameters of each frame, all of them have the same format as the Cape dataset and are aligned well. So I'm wondering what the problem is? Do I need just more epochs in the first stage, or something else? I hope you can give me some advice, thank you!

[shunsukesaito]

I have not had a chance to test the code on ClothSeq dataset. My guess is either global scale or vertex density of input scans is different, potentially requiring different hyper parameters.

[Bill-WangJiLong]

我没有机会在ClothSeq数据集上测试代码。我的猜测是输入扫描的全局尺度或顶点密度不同，可能需要不同的超参数。

How should the global scale be modified to suit?