Paper Overview
Field: Computer Vision Authors: Iván Verdugo Guerra, Ezequiel López Rubio, Jorge García González arXiv: 2610.08756
Summary
The same Gaussian of a 3D Gaussian Splatting model is seen from many views, and these views do not always agree on the class it belongs to. The Gaussian may be occluded in some views, and the detector's confidence differs from one view to another. Ground truth is given as an annotated mesh, since two training runs do not produce identical Gaussians.
This work proposes a post-training lifting method that handles one target class at a time and combines information from all views. Target and non-target evidence are accumulated simultaneously, weighted by the visibility of each Gaussian in each view. The Gaussians are then filtered with two thresholds: a main threshold β selects high-confidence seeds, and a lower threshold γβ adds connected components around them. During evaluation, labels are transferred from Gaussians to mesh vertices that are both visible and annotated.
This design separates three error sources: the 2D detector, the lifting step, and the transfer between representations.
Results
- Validation scenes: mean mIoU of 0.93 using dataset-annotated masks, 0.65 using YOLO masks
- ScanNet++ test scenes: 0.80 and 0.54 respectively
- Compared to a previous version that thresholded per-view evidence, the improvement raises test mIoU by 0.24 and enables a single threshold across all classes and scenes of both datasets
- Error analysis shows most remaining errors originate from the detector
Original Abstract (excerpt)
> The same Gaussian of a 3D Gaussian Splatting model is seen from many views, and these views do not always agree on the class it belongs to. The Gaussian may be occluded in some of them, and the confidence of the detector is not the same from one view to another. The ground truth, on the other hand, is given as an annotated mesh, because two training runs do not produce the same Gaussians. In this work, we propose a post-training lifting method that works with one target class at a time and combines the information coming from all the views. Target and non-target evidence are accumulated simultaneously, weighted by the visibility of each Gaussian in each view. After that, the Gaussians are filtered with two thresholds: a main threshold β selects the high-confidence seeds, and a lower one γβ...
---
*Auto-collected on 2026-10-08*