demo mesh result is different

[yanqi1811]

Thanks for your work!
When I run the demo using the image.jpg, I get the different result with you! The bbox and 2d points skeleton are same with you, but the mesh is different. I also test other images, I find it doses't even work well, when fingers down. I'm not sure what the problem is. Could you help me?

[mks0601]

which checkpoint did you use?

[yanqi1811]

which checkpoint did you use?

This pre-trained InterWild! Download from here.

[mks0601]

I've just ran the demo and I got this result from that checkpoint

Also, I got this result from a checkpoint trained on ih26m+mscoco+reih

[yanqi1811]

I've just ran the demo and I got this result from that checkpoint

Also, I got this result from a checkpoint trained on ih26m+mscoco+reih

Thank you very much. I've just ran the demo and I got this result from the pretrain checkpoint.

I also got this result from a checkpoint trained on ih26m+mscoco+reih.

But these results are different from you. Is there something wrong with my environment, like smplx?

[mks0601]

that's weird. I'm using original version of smplx. Nothing is modified on my side. are you using customized smplx?

[yanqi1811]

that's weird. I'm using original version of smplx. Nothing is modified on my side. are you using customized smplx?

I am also using original version of smplx, the version is 0.1.28. What about you?

[mks0601]

I think it's not a problem of smplx.. are you using my code as it is without modification? are you using pytorch3d?

[yanqi1811]

I think it's not a problem of smplx.. are you using my code as it is without modification? are you using pytorch3d?
It is not any modification. Yes, I am using pytorch3d when I render mesh result.

[mks0601]

It seems the root pose is flipped 180 degree.
Could you check your pytorch3d version with
import pytorch3d
print(pytorch3d.version)

I'm using 0.7.3

[yanqi1811]

It seems the root pose is flipped 180 degree. Could you check your pytorch3d version with import pytorch3d print(pytorch3d.version)

I'm using 0.7.3

I'm using 0.7.4. I will try pytorch3d 0.7.3. Thanks.

[yanqi1811]

It seems the root pose is flipped 180 degree. Could you check your pytorch3d version with import pytorch3d print(pytorch3d.version)
I'm using 0.7.3

I'm using 0.7.4. I will try pytorch3d 0.7.3. Thanks.

I use pytorch3d 0.7.3, the result is right. Thanks!

[mks0601]

Oh I didn't know that pytorch3d is very sensitive to versions

[yanqi1811]

Oh I didn't know that pytorch3d is very sensitive to versions
It seems like I've found the real cause of the problem. When converting the ONNX file earlier, I discovered that the aten::atan2 operator is not supported. I took a bold step and modified the implementation of torch.atan2() in the PyTorch3D source code by replacing it with torch.atan(). Specifically, I replaced atan2(A, B) with atan(A / (B+1e-6)). In my simple tests, I didn't notice any significant impact on the upward finger direction. However, it appears that the impact is actually quite significant. Do you have any alternative replacement suggestions for the torch.atan2() function that would ensure ONNX compatibility? Your advice would be greatly appreciated. Thank you.
This pytorch3d original code:

[mks0601]

Hi, sorry I haven't used ONNX..

[yanqi1811]

Hi, sorry I haven't used ONNX..

OK, thanks.