search

zju3dv / PGSR

code for "PGSR: Planar-based Gaussian Splatting for Efficient and High-Fidelity Surface Reconstruction"
https://zju3dv.github.io/pgsr/
Other
433 stars 25 forks source link

Multi-camera data #17

Open li1u opened 1 month ago

li1u commented 1 month ago

Hello, I am amazed by the results of your paper, it is really beautiful. But now there is a problem, when I process multi-camera data (renderme360 dataset), the effect becomes strange. image

I have adjusted the code for the data reading part (json): image I don't know why, but if you read cx and cy directly, you will get an error, so use prcppoint to calculate. projection matrix:

image image

The projection matrix in sample_cam has also been modified accordingly.

30000_00000014 (The last image)

danpeng2 commented 1 month ago

Hello, thank you for recognizing our work.

We have not yet conducted a complete test on the configuration where the camera intrinsic parameters are not at the center of the image. I suggest cropping the original image so that the camera intrinsic parameters are at the center of the image.

li1u commented 1 month ago

image It should just be that the size is set too large, but I feel there are still some problems that need to be solved.

danpeng2 commented 1 month ago

The surface noise is quite significant. If there are no special reflections or lighting changes in the image, the surface should be relatively smooth, indicating there are indeed some issues that need to be addressed. Here are my recommended troubleshooting directions:

  1. Check the output multi_view.json to confirm if appropriate frames have been selected.
  2. Ensure the camera poses are accurate enough. It is best to use COLMAP for fine-tuning.
  3. Avoid using unverified features like getProjectionMatrixCenterShift.