Frosting
Frosting copied to clipboard
some question about reconstruction result
thank you for your great work! i have several questions about reconstructed mesh:
- I filmed a video indoors with my Android phone, 1080p, sample about 160 images, reconstructed it with COLMAP. The normal direction of the generated mesh is reversed. Is this due to the difference in coordinate systems?
- I want to do a high-fidelity reconstruction of the object in the middle and found that the detail modeling effect is not very good. Are there any methods to enhance the details?
- if i replace vanilla_gs with other gs method, such as Gaussian Pro or 2DGS, will the reconstruction results be better?
- If I use the parameters bboxmin and bboxmax to limit the reconstruction range, will I get better results?
- I have tried all the regularization types, and it seems that SDF (Signed Distance Field) provides the best results.
Additional note: I tried using the PLY results from 7,000 iterations of Gaussian splatting and 30,000 iterations of Gaussian splatting. The 30,000 iterations gave me a better result.
Here are some screenshots.
video images:
mesh normal direction is not correct:
object i want to reconstruct:
gaussiang splatting 7000 iterations:
gaussiang splatting 30000 iterations: