GVC-Seg: Training Free 3D Instance Segmentation
via Geometric Visual Correspondence

Liang Xu1, Fangjing Wang1, Jinyu Yang2,
Feng Zheng1,
1Southern University of Science and Technology 2tapall.ai
Paper Code

We present GVC-Seg , a novel training-free 3D instance segmentation framework. By combining 3D geometric and 2D visual cues in a training-free manner, GVC-Seg prompts more reliable proposal generation and selection, which alleviates the confidence bias of multi-scale, multi-source proposals.

Summary of results using GVC-Seg

Summary of results using GVC-Seg on ScanNet200, ScanNet++ and Replica dataset.

Visualizing results for GVC-Seg

Video Visualization.

Geometric Visual Correspondence

We show a pair of predicted instance masks and confidence scores.

Qualitative results of different methods on four instances.

The quality of our masks is clearly superior to others.

BibTeX


      @article{
        author    = {Liang Xu†, Fangjing Wang†, Jinyu Yang, Feng Zheng},
        title     = {GVC-Seg: Training Free 3D Instance Segmentation via  Geometric Visual Correspondence.},
        journal   = {},
        year      = {2024},
      },