Emil Bogomolov

@nicolas-chaulet is it possible to decompose this on tasks? I will contribute

Can't we just use segmentation tool from ScanNet repository for this?

> > Can't we just use segmentation tool from ScanNet repository for this? > > Do you mean the [segmentator](https://github.com/ScanNet/ScanNet/tree/master/Segmentator)? It sounds good. And can it be used for other...

Why don't you use a provided docker?

I am completely agree with you that will make training slower. My collegue and me concluded that your approach also leads to convergence. But I wish there would be thoughts...

@jinfagang yes, please could you share this knowledge. At least in words

I'm also encountered this question. One thing that I've deduced is whenever we use VIBE or PARE the `pred_cameras` are needed to be converted as it done in `visualize_smpl_vibe` flow....

Regarding the location. Have you tried `Vector((100*trans[1], 0, 100*trans[0]))`.

Yeah, you right it is a [N,3] tensor, but consider that these parameters are weakperspective camera parameters not an x,y,z. Follow `visualize_smpl_vibe` flow to get `orig_camera` or `K, R, T`...

I could achieve the following performance. 3D model is well aligned but whenever the person on the video do "sit" the model do "jump" [tiktok source](https://www.tiktok.com/@matayasweeting/video/7045393332893404463?is_from_webapp=v1&item_id=7045393332893404463) https://user-images.githubusercontent.com/25350960/186831431-16f886b5-6d6a-4b27-8fa9-3034bc8e0e52.mp4