ShichenLiu / SoftRas

Project page of paper "Soft Rasterizer: A Differentiable Renderer for Image-based 3D Reasoning"

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

how to build and train our own dataset ?

M202071895 opened this issue · comments

I have an obj model, how to build a dataset like him (a set of 24 pictures),using 3Dsmax ? How to take this group of pictures ?

You only need to make sure that your rendering setting (extrinsic/intrinsic/look at) is the same as that of the SoftRas.

You only need to make sure that your rendering setting (extrinsic/intrinsic/look at) is the same as that of the SoftRas.

thank you for your reply !It's hard for me to do this, could you please share some more details?When I use 24 cameras and render these pictures in 3dsmax, The result is always wrong and can't train the models.
image

Hi, the camera position seems correct. I assume they are all with 30degree of elevation. Can you show me some of your rendered images. Have you tried to compare your renderings with that in the dataset, e.g. same id with similar appearance? I have no idea about what you meant “wrong”.

Sorry I couldn't find your images. Do you mind uploading them to a public storage, e.g. google drive?

Sorry I couldn't find your images. Do you mind uploading them to a public storage, e.g. google drive?

It's my dataset:

Sorry I couldn't find your images. Do you mind uploading them to a public storage, e.g. google drive?

It's my dataset:
X (~6Y(XG@}9K9FWFT$1EY

and It's yours:

ZU)F)U0OKAZMM87(8}~17EM

This is weird, can you send me the numpy you used? I want to try it on my side.

This is weird, can you send me the numpy you used? I want to try it on my side.

Can I communicate with you by email, It will be better.

The rendering process in the current code does not include textures, and the reconstruction basically only considers contours and shapes. How can I add the texture of the model, and whether the reconstruction effect will be better after adding it.