search

CUHK-AIM-Group / EndoGaussian

EndoGaussian: Real-time Gaussian Splatting for Dynamic Endoscopic Scene Reconstruction
https://yifliu3.github.io/EndoGaussian/
MIT License
115 stars 6 forks source link

Dynamic Modeling on Scared Dataset #30

Closed haowang020110 closed 2 months ago

haowang020110 commented 2 months ago

Hi, there It seemed that the Scared Dataset is a static scene and I noticed that you keep modeling dx , do (didn't modeling ds, dq) in pratice. https://github.com/CUHK-AIM-Group/EndoGaussian/blob/47c23a953fa211835a3d4113f0075e7815bebeca/arguments/scared/d1k1.py#L40-L43 I try to set them all to True to train (i.e. wish to use the vanlia 3dgs). The results get worse and the render image was blur. I guess the bad results arise from the inaccurate pose and depth. Could you give me some insight about the reason of your setting and the negative factor in the data.

haowang020110 commented 2 months ago

Thanks a lot for your work and sorry for any inconvenience caused!

yifliu3 commented 2 months ago

Hi, there Yes the reason may come from inaccurate camera poses and depths, and may also come from moving lights attached to the endoscope. If you want to do research only on the dynamic setting, you can try hamlyn or stereomis dataset.

haowang020110 commented 2 months ago

Thanks a lot for your fast reply!