search

BaowenZ / RaDe-GS

RaDe-GS: Rasterizing Depth in Gaussian Splatting
Other
460 stars 25 forks source link

Poor training results #15

Closed hefan14 closed 3 months ago

hefan14 commented 3 months ago

I reproduced the paper and trained it according to the README.md, but the results of 7000 and 30000 iterations were very poor. I used Meshlab to visualize the training results point_cloud.ply. For the training results of scan65 in the DTU dataset, it is almost impossible to see that it is a skull. Could you help me solve this?

BaowenZ commented 3 months ago

Wrong glm version will make the training not converge. Please check whether the glm is in the specific version, and whether you use the latest version of our code.

hefan14 commented 3 months ago

Thanks! I will recheck and modify it.

hefan14 commented 3 months ago

I downloaded the code using the 'git clone' command, which includes glm@5c46b9c and retrained. The training results were the same as before(What I'm referring to is that the point_cloud at 7000 and 30000 steps cannot see the outline of the dataset image. May I ask if this situation occurs during your training). May I ask if my GLM version is correct？