Less Context, Better Geometry: Masked Geometric Encoder for Robust 3D Foundation Models
Zhimin Shao, Xijun Liu, Zhaoliang Zhang, Yutao Tang, Abhay Yadav, Rama Chellappa, Cheng Peng
News
- [2026/10/2] Training/evaluation code release.
Overview
Given a sequence of images, during training, MGE masks the global attention randomly to limit the all-to-all cross-view information flow. To promote richer intermediate feature representation, we further introduce constrain MGE to produce similar 3D features compared to the full-attention teacher encoder output.
Reconstruction Visualizations
Installation
- Clone MGE
git clone https://github.com/Zhimin00/MGE.git
cd MGE
- Create conda environment
conda env create -f mge.yaml
Download Checkpoints
Please download pretrained teacher model from here and save at ckpt/.
The checkpoint of MGE is available at both Hugging Face. Download it and save at checkpoints/.
Citation
If you find our work useful, please consider citing: