Research map / 3D vision / 3D generation
What VGGT Knows About Overlap: Probing Geometric Foundation Models for Co-Visibility
Filippo Ziliotto, Luciano Serafini, Lamberto Ballan, Tommaso Campari
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train layer-wise MoE head on frozen VGGT (vs zero-shot) | Co-VGGT · Gibson pairwise | Graph IoU*: 0.50 -> 0.85 | Improved |
| train layer-wise MoE head on frozen VGGT (vs zero-shot) | Co-VGGT · Gibson multiview | Graph IoU*: 0.37 -> 0.74 | Improved |
| add |e_i-e_j| to pair features (ablation) | Co-VGGT aggregator · Gibson multiview | Graph IoU*: 0.71 -> 0.74 | Improved |
| add e_i*e_j product to pair features (ablation) | Co-VGGT aggregator · Gibson multiview | AUC: 0.71 -> 0.73 | Improved |
| add product feature to simple concatenation (ablation) | Co-VGGT aggregator · Gibson multiview | Graph IoU*: 0.67 -> 0.71 | Improved |
| add layer-wise MoE over standard aggregation (ablation) | Co-VGGT · Gibson multiview | Graph IoU*: 0.69 -> 0.74 | Improved |
| add layer-wise MoE over standard aggregation (ablation) | Co-VGGT · HM3D multiview | Graph IoU*: 0.72 -> 0.76 | Improved |
| use all layers [1,24] instead of last 12 layers [14,24] (ablation) | Co-VGGT MoE · Gibson multiview | AUC: 0.72 -> 0.73 | Improved |
| use all layers [1,24] instead of only last layer [24] (ablation) | Co-VGGT MoE · Gibson multiview | Graph IoU*: 0.69 -> 0.74 | Improved |
| use all layers instead of only first 12 layers [1,12] (ablation) | Co-VGGT MoE · Gibson pairwise | Graph IoU*: 0.30 -> 0.84 | Improved |
| use last 5 layers [19,24] for zero-shot similarity instead of all layers (ablation) | Co-VGGT zero-shot · Gibson pairwise | Graph IoU*: 0.50 -> 0.52 | Improved |
| train on Gibson, test on HM3D (cross-domain) | Co-VGGT · HM3D pairwise | Graph IoU*: 0.84 -> 0.81 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 4 of 4 on most shared results.
- DUSt3R — DUSt3R: Geometric 3D Vision Made Easy (CVPR 2024) · better on 2 of 2 shared results
- CroCo v2 — CroCo v2: Improved Cross-view Completion Pre-training for Stereo Matching and Optical Flow (ICCV 2023) · better on 2 of 2 shared results
- MV-DUSt3R+ — MV-DUSt3R+: Single-Stage Scene Reconstruction from Sparse Views In 2 Seconds (CVPR 2025) · better on 1 of 1 shared result
- SuperGlue — SuperGlue: Learning Feature Matching With Graph Neural Networks (CVPR 2020) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- LiteVGGT: Boosting Vanilla VGGT via Geometry-aware Cached Token Merging (CVPR 2026 · similarity 0.88)
- VGGT: Visual Geometry Grounded Transformer (CVPR 2025 · similarity 0.87)
- Feat2GS: Probing Visual Foundation Models with Gaussian Splatting (CVPR 2025 · similarity 0.86)
- World Tracing: Pixel-Aligned Geometry Beyond the Visible (NeurIPS 2026 · similarity 0.86)
- FrameVGGT: Coherence-Preserving Memory for Bounded Streaming Geometry (NeurIPS 2026 · similarity 0.86)
- Probing the 3D Awareness of Visual Foundation Models (CVPR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
