Compared to other algorithms, there are more noise and fuzzy areas

[CanCanZeng]

Hello, this job is very innovative. Thank you for opening up this code!
I found that this algorithm can solve the problem of large areas of blurring in very messy places in the original 3DGS algorithm. The usual solution to this problem is to use AbsGS, but I have found that the MCMC algorithm can also solve this problem.
But I also found that the MCMC algorithm has some unique phenomena.
For example, the generated point cloud is noiser compared to other algorithms, and there will be a lot of distant noise, which is rarely seen in other algorithms. I guess the reason for this phenomenon is that the algorithm adds positional noise during each iteration, causing some GS positions to constantly move randomly and reach very far places. Although the author used transparency and size constraints to partially suppress this situation, it cannot completely solve it.
Additionally, in certain areas, there may be more blurriness, typically with tree branches in the sky as the background. I guess this phenomenon is caused by the sampling mechanism. There are fewer point clouds in the branch area, and it is likely that the opacity is low at the beginning, so the probability of this area being sampled is lower. During the iteration process, as the reconstruction area of the core becomes more point clouds, the proportion of point clouds in the branch area is less, making it more difficult to be sampled. The final result is that the number of point clouds in this area is less, and the rendering result is that this part is relatively blurry.
I'm not sure if my understanding is correct, but does the author have any good suggestions for improving these two issues?

Pay attention to the upper left corner of the image below

[shakibakh]

@CanCanZeng Thank you for your interest in our work.

In our formulation, we developed a general probabilistic framework. We do add unconstrained noise to the positions of the Gaussians, which can result in them moving to distant locations, as you mentioned. If any positional constraints are known a priori, incorporating them could be beneficial; however, in our approach, we assumed no prior information.

Regarding your second issue, an error-based Gaussian relocation technique might be helpful. There are several papers that explore error-based sampling, which you might find useful.

[CanCanZeng]

Thank you for your answer. Could you please indicate the name of the paper or algorithm you mentioned?

[FunnyPocketBook]

@CanCanZeng The MCMC paper mentions Revising Densification in Gaussian Splatting
as one of the error-based sampling methods. So far I have not found an existing implementation of this paper