Yan Tian · Weiping Ding · Mahmoud Hassaballah · Karen Egiazarian
Aura Conci · Abdulkadir Sengur · Leszek Rutkowski
Paper · Project page source · Contact
MGMA combines hierarchical feature distillation (HFD) and multimodal semantic alignment (MSA) in a shared 3D Gaussian representation.
- HFD maintains separate semantic and hierarchical feature streams. SAM-derived regions provide supervision for semantic contrastive learning and hierarchical feature learning.
- MSA applies category-text supervision to both streams at the Gaussian and rendered-image levels, forming four alignment pathways.
- Open-vocabulary evaluation expands the text bank from seen categories to the joint seen–unseen vocabulary.
The manuscript abstract reports the following mIoU improvements over OpenGaussian:
| Dataset | Evaluation categories | Seen / unseen | mIoU improvement |
|---|---|---|---|
| Shining3D | 16 | 11 / 5 | +5.05 percentage points |
| Aoralscan3 | 17 | 11 / 6 | +5.16 percentage points |
Left to right: RGB image, ground truth, OpenGaussian baseline, +HFD, +MSA, and +HFD & MSA. Black boxes highlight comparison regions. Reproduced from Figure 5 of the supplied manuscript.
Title, abstract, authors, and figures are based on the supplied manuscript. The supplied copy contains eight pages; the numerical improvements above come from its abstract. The research task and title follow the PDF body, which describes semantic segmentation.
The page layout is inspired by SSC-BSS-OVSS. This repository contains the project website and its assets. The supplied materials do not include the model implementation, model checkpoints, or public arXiv and DOI links.
See 使用说明 for local preview, GitHub Pages setup, and image regeneration.