Jumpat
commented
1 day ago Hi, I wonder what kind of noise you want to get rid of. There are some floaters in the space due to the inherent defect of 3D-GS. And there may be some segmentation failure due to the defect of SAGA. Another issue is the clustering algorithm. We randomly select some Gaussians to conduct clustering and use the cluster centroids to calculate their average features as query features. Then these query features are used to classify each 3D Gaussian. You may need to check this part for modification.
Hello, everyone! I ran the reconstruction on figurines dataset and I like the way it segmented table and all the objects on it, but there is too much noise around it. Is there any way to handle that noise? For example something like a confidence level, where noise could have low confidence to easily get rid of it or some separate background mask. Or any other way?