hello，have you ever run your method on initialized sfm point cloud by colmap？
Closed this issue · 8 comments
hello，have you ever run your method on initialized sfm point cloud by colmap？
Hello，I have test your method using initialized sfm point cloud by colmap，but I found that the effect of your method is worse than 3dgs ... Is there something wrong with the code implementation?
Hi, what is the "initialized sfm point cloud by colmap" referred to? In our test, COLMAP cannot estimate correct point clouds for most of the scenes only using 3 valid views especially in DTU and Blender, so this situation is not in our consideration. If you mean the point clouds that are estimated by all views, it certainly could be better for 3DGS since extra priors from unseen views are provided, but it is unreasonable for few-shot setting.
There is another possible initialization that uses the MVS point cloud like what FSGS did. However, we have not adapted our code to this situation yet.
Hello，I have test your method using initialized sfm point cloud by colmap，but I found that the effect of your method is worse than 3dgs ... Is there something wrong with the code implementation?
Hi, what is the "initialized sfm point cloud by colmap" referred to? In our test, COLMAP cannot estimate correct point clouds for most of the scenes only using 3 valid views especially in DTU and Blender, so this situation is not in our consideration. If you mean the point clouds that are estimated by all views, it certainly could be better for 3DGS since extra priors from unseen views are provided, but it is unreasonable for few-shot setting.
There is another possible initialization that uses the MVS point cloud like what FSGS did. However, we have not adapted our code to this situation yet.
Hello，I have test your method using initialized sfm point cloud by colmap，but I found that the effect of your method is worse than 3dgs ... Is there something wrong with the code implementation?
I don't understand what you mean by "not adapted to". In fact, I only used 3 training perspective images to reconstruct point clouds using colmap. The processing method is the same as FSGS. Under the same point cloud initialization, number of training rounds and various conditions, your method generates The new perspective is full of fragmented artifacts and floating points, which is about 1 lower than the PSNR value of 3dgs. At the same time, the visual effect is far worse than 3dgs, let alone fsgs. In other words, your method is just a little better than 3dgs in the case of chaotic initialization of Gaussian, but in fact it only takes tens of seconds for colmap to perform the mvs operation on three images. I still think the code you gave may be some problems. 😂
There is a problem with the video display of fsgs. You can directly watch the demo video they gave.
https://github.com/zhiwenfan/zhiwenfan.github.io/blob/master/Homepage_files/videos/FSGS_gif.gif
I have also recognized that MVS initialization can bring huge improvements even when only using the original 3DGS. That's also why we recommend using it in the README.
"not adapted to" means the current hyperparameters are set for the random initialization situation. If such a much stronger prior is in use, I think we should at least adjust some hyperparameters to reduce the effect of regularization terms, e.g. position_lr, or a smaller weight of regularization terms, just like FSGS that also uses a very small weight for its depth loss term. It's obviously meaningless if you directly use the config we set for our setting to evaluate the method on another initialization...
hh thanks
Thanks for your interest! 😊