GaussianWorld/scannetpp_v2_default_fixed_xyz_3dgs

Please note for this version of ScanNet++ v2 3DGS scenes, we disabled the optimization of 3DGS centers and used exactly the same number and locations as the initializing point clouds, so we can directly copy the original scene sem.seg. labels to the GS. This did cause strong limitation on the resulted 3DGS quality and this part of the data are not used for SceneSplat vision-language pretraining, but only for self-supervised training and later SceneSplat main paper Table 4 experiments of supervised semantic segmentation with 3DGS inputs.