Yonsei University's VODA Fixes 3D Object Editing by Supervising Hidden Geometry
VODA shows that 3D Gaussian Splatting segmentations with clean 2D masks often hide incomplete object selections, and fixes it with depth-cluster supervision.
Takeaways
- VODA is a plug-and-play loss fixing incomplete 3D Gaussian selections hidden behind accurate rendered masks. Paper
- It groups per-pixel Gaussians by depth gaps and supervises every member of the selected cluster directly.
- New Removal-3D benchmark and mRDE metric score depth change after deleting selected Gaussians.
- LaGa + VODA cuts mRDE from 36.93 to 21.66 and lifts LERF mIoU from 56.80 to 60.28.
- All four tested baselines improve on Removal-3D while holding or gaining on LERF and ScanNet.
- Code at github.com/jhxcho/voda; project page with interactive demos.
A new NeurIPS 2026 paper from Yonsei University identifies a hidden failure in 3D Gaussian Splatting segmentation. Existing methods can render crisp 2D masks while selecting little more than an object’s visible shell. Delete the Gaussians labeled as a sheep, for example, and much of the sheep can remain in the scene.
The researchers propose VODA, a training objective that extends supervision through depth clusters containing both visible and occluded Gaussians. It integrates with existing segmentation pipelines while leaving their architectures and inference procedures unchanged. More complete selections improve object removal, recoloring, scaling, and repositioning.
A crisp mask can hide a hollow selection
3D Gaussian Splatting represents a scene with millions of semi-transparent Gaussian primitives. A renderer blends them from front to back to produce an image. Segmentation methods attach features or labels to each Gaussian, allowing applications to select and edit objects.
Training usually compares a rendered feature map with a 2D segmentation mask and backpropagates the error. Each Gaussian’s gradient depends on its opacity and transmittance, the amount of light that remains after the renderer processes primitives closer to the camera. Gaussians behind an opaque surface contribute little to the pixel, so they receive little supervision.