search

ShichenLiu / SoftRas

Project page of paper "Soft Rasterizer: A Differentiable Renderer for Image-based 3D Reasoning"
MIT License
1.24k stars 156 forks source link

how to build and train our own dataset ？ #82

Open M202071895 opened 3 years ago

M202071895 commented 3 years ago

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 ？

ShichenLiu commented 3 years ago

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

M202071895 commented 3 years ago

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

ShichenLiu commented 3 years ago

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”.

ShichenLiu commented 3 years ago

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

M202071895 commented 3 years ago

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

ShichenLiu commented 3 years ago

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

M202071895 commented 3 years ago

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.

M202071895 commented 3 years ago

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.