Research map / Vision-language and multimodal / Human and embodied modeling
SceneGraphLoc: Cross-Modal Coarse Visual Localization on 3D Scene Graphs
Yang Miao, Francis Engelmann, Olga Vysotska, Federico Tombari, Marc Pollefeys, Daniel Barath
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 |
|---|---|---|---|
| use SceneGraphLoc scene-graph embeddings instead of LidarCLIP / LIP-Loc point-cloud global descriptors | cross-modal coarse localization · 3RScan test, 50 scenes, static | Recall@1: 4.7 (LidarCLIP) -> 30.2 | Improved |
| add images to the scene graph map embeddings (ablation) | SceneGraphLoc · 3RScan test, 50 scenes, temporal | Recall@1: 28.2 -> 69.3 | Improved |
| add attributes to point cloud map modality (ablation) | SceneGraphLoc · 3RScan val (temporal) | Recall@1: 43.9 -> 54.8 | Improved |
| add relationships modality (ablation) | SceneGraphLoc · 3RScan val (temporal) | Recall@1: 56.5 -> 62.7 | Improved |
| add structure modality on top of point cloud and attributes (ablation) | SceneGraphLoc · 3RScan val (temporal) | Recall@1: 54.8 -> 56.5 | Improved |
| add attributes, structure, relationships to point cloud plus images (ablation) | SceneGraphLoc with GCViT features · 3RScan val (temporal) | Recall@1: 73.2 -> 72.1 | No clear effect |
| use DINOv2 instead of GCViT image features (ablation) | SceneGraphLoc · 3RScan val (temporal) | Recall@1: 72.1 -> 88.5 | Improved |
| increase number of views per object beyond 10 (ablation) | image modality embedding · 3RScan val (temporal) | Recall@1: 88.5 (10 views) -> 86.3 (15) / 86.8 (20) | No clear effect |
| use transformer encoder with pose positional encoding and max-pooling for multi-view fusion (ablation) | image modality embedding · 3RScan val (temporal) | Recall@1: 85.5 -> 88.5 | Improved |
| use SceneGraphLoc without images versus OpenMask3D on ScanNet all scenes | scene graph map without images · ScanNet test (all 210 scenes) | Recall@1: 16.5 (OpenMask3D) vs 13.5 | Hurt |
Closest papers by meaning
- CrossOver: 3D Scene Cross-Modal Alignment (CVPR 2025 · similarity 0.85)
- Universal Scene Graph Generation (CVPR 2025 · similarity 0.85)
- Multiview Scene Graph (NeurIPS 2024 · similarity 0.84)
- Cross-Modal Distillation for 2D/3D Multi-Object Discovery from 2D Motion (CVPR 2025 · similarity 0.84)
- Object-X: Learning to Reconstruct Multi-Modal 3D Object Representations (NeurIPS 2025 · similarity 0.84)
- CLIP-Driven Open-Vocabulary 3D Scene Graph Generation via Cross-Modality Contrastive Learning (CVPR 2024 · similarity 0.84)
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.
