Why rasterization process is so low, how to speed it.

[y6216886]

Thank you for your time!
I have tried to add some noise to the smplx model by setting noise_scale in the yaml file to [0.1, 0.1].
However, it requires to rasterize meshes based on the perturbed smplx model, which leads to running function rasterize_meshes() in line 353 in lib/dataset/mesh_util.py.
pix_to_face, zbuf, bary_coords, dists = rasterize_meshes( meshes_screen, image_size=raster_settings.image_size, blur_radius=raster_settings.blur_radius, faces_per_pixel= raster_settings.faces_per_pixel, bin_size=raster_settings.bin_size, max_faces_per_bin= raster_settings.max_faces_per_bin, perspective_correct=raster_settings.perspective_correct, cull_backfaces=raster_settings.cull_backfaces, )
This function runs very slow in my host (40 minutes for an iteration), any idea to speed this?

PS: I have tried setting parameters bin_size to none, or setting face_per_pixel to 1, but these operations cannot work.

Thanks.

[y6216886]

the reason seems to be the large value of the parameter "image_size" of the Pytorch3dRasterizer.
rasterizer = Pytorch3dRasterizer(image_size=2**12)
Also, the rasterization process is running on CPU mode, which further slows down the speed.

So, I try to set the parameter image_size from "2^12" to "2^9" and the rasterization process now get much faster,
But, will this operation hurt performance or is it a wrong operation?

[YuliangXiu]

I don't think this will hurt the performance a lot, you can visualize the pre-computed visibility value (https://github.com/YuliangXiu/ICON/blob/master/docs/dataset.md#generate-mode) on SMPL-X obj to check if it works well.

Also, please use GPU version, otherwise the speed is unacceptable.