Tianfang-Sun
Tianfang-Sun
> Hi, thank you for your question. One possible reason for the fake points might be the spurious estimation of stereo matching. Our accumulated point cloud is composed of observations...
> Hello @isunLt > > I am using also the accumulated point-clouds alongside the RGB frames but I faced another issue actually which is some files are empty. For example...
The README is confusing. It says that files in datasets/Cityscapes/rgb/train is expected to be png files. But the code expects it to be .ppm files. And the pretrained weights also...
> Does any one know this same problem? I think "PPM" file format change to jpg and Datasets directory structures must be recontructed. If I changed ppm to png, The...
Same error
不好意思打扰了，我把python换成3.7，pytorch换成1.6之后就没问题了。
抱歉我又来了，在换成python3.7、pytorch1.6以后，到训练的最后还是出现了老问题。 
> 还有种可能是要在所有normalize的分母处加一个 1e-9 或者 1e-12。因为不知道什么原因分母的norm值训练的太小了，但是我自己没遇到类似问题。 > >  谢谢您，我去试一下。
> 请问您问题解决了吗？我改了norm还是会出现nan，我的loss下降很快，在一个epoch内就变nan了 太久了，我忘记了，不好意思
@soulslicer if you have a [N, 6] matrix, you need to reshape the matrix to [N, 4] of form (x, y, z, rgb). You can see the way they express...