lukasHoel / text2room

Text2Room generates textured 3D meshes from a given text prompt using 2D text-to-image models (ICCV2023).
https://lukashoel.github.io/text-to-room/
MIT License
1.02k stars 71 forks source link

Results got worse after quadratic decimation and poisson #21

Closed Jihyun0510 closed 1 year ago

Jihyun0510 commented 1 year ago

I always appreciate your kind answers and your great work

Running the codes (generate_scene.py) with the default setting, I found out that results get worse as time passes.

Here are the results for the living_room.json:

Fused frame until 200 image

Fused final image

After poisson image

*After qudratic decimation image

Could you explain the reasons? What should I do to get the best results like in your paper?

lukasHoel commented 1 year ago

Hi you can take a look at the discussion in #13 and try out those steps. If the ceiling looks like this after the first stage, the default parameters for completion are unable to inpaint it properly. That is, our camera sampling fills in some holes, but others remain. As a result, poisson surface reconstruction will create these smoothing artifacts to make the mesh watertight. Also see Fig9. in the paper for a discussion about this failure mode of the method.

The two solutions are:

Jihyun0510 commented 1 year ago

Thank you! I will try that