Research map / 3D vision / Depth and stereo
PGGT-Loc: Primitive-Grounded Geometry Transformer for Feed-Forward Camera Localization
Hanqiao Ye, Yuzhou Liu, Mengqi Rong, Jian Liu, Yangdong Liu, Shuhan Shen
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PGGT-Loc, a primitive-grounded geometry transformer for feed-forward camera localization that uses compact scene primitives from readily available structured maps instead of dedicated costly maps.
Problem
Camera localization typically relies on multi-stage pipelines and dedicated maps that are costly to construct, store, and maintain, while well-structured primitive-based maps that are often available have not been exploited for direct camera localization.
Approach
Uses compact, interpretable scene primitives from existing structured maps as the localization reference, grounding a geometry transformer in these primitives to regress camera pose feed-forward rather than through a multi-stage pipeline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add patch-primitive assignment loss (ablation) | PGGT-Loc (Ray) · ScanNet | pose recall (fine threshold, %): 54.7 -> 62.0 | Improved |
| add primitive parameter loss (ablation) | PGGT-Loc (Ray) · ScanNet | pose recall (fine threshold, %): 54.9 -> 62.0 | Improved |
| add prepended camera/map token pose loss (ablation) | PGGT-Loc (Ray) · ScanNet | pose recall (fine threshold, %): 59.7 -> 62.0 | Mixed |
| add semantics injection alone (ablation) | PGGT-Loc layout · S3D-resampled 90 deg FoV | 1 m median translation error (cm): 15.2 -> 17.3 | Hurt |
| add room isolation attention mask alone (ablation) | PGGT-Loc layout · S3D-resampled 90 deg FoV | pose recall at 1 m (%): 90.5 -> 93.8 | Improved |
| add semantics and isolation jointly (ablation) | PGGT-Loc layout · S3D-resampled 90 deg FoV | pose recall at 1 m (%): 90.5 -> 95.5 | Improved |
| use full multi-task supervision instead of ray loss only (ablation) | PGGT-Loc (Ray) · ScanNet | pose recall (fine threshold, %): 51.7 -> 62.0 | Improved |
| derive pose from ray map instead of plane parameters (ablation) | PGGT-Loc · ScanNet | pose recall (fine threshold, %): 59.5 -> 69.0 | Improved |
| derive pose from ray map instead of plane parameters (ablation) | PGGT-Loc layout · S3D-official | pose recall at 1 m (%): 76.7 -> 93.6 | Improved |
| use camera token instead of ray map (ablation) | PGGT-Loc · ScanNet | pose recall (fine threshold, %): 69.0 -> 68.9 | No clear effect |
| remove plane parameter inputs and prediction (ablation) | PGGT-Loc · ScanNet | coarse recall | No clear effect |
| feed-forward grounding instead of render-and-compare | PGGT-Loc vs SPVLoc · S3D-official | pose recall at 1 m (%): 84.5 -> 93.6 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- PlanaReLoc — PlanaReLoc: Camera Relocalization in 3D Planar Primitives via Region-Based Structure Matching (CVPR 2026) · better on 7 of 7 shared results
- PlaneRecTR++ — PlaneRecTR: Unified Query Learning for 3D Plane Recovery from a Single View (ICCV 2023) · better on 4 of 4 shared results
- SPVLoc — SPVLoc: Semantic Panoramic Viewport Matching for 6D Camera Localization in Unseen Environments (ECCV 2024) · better on 3 of 3 shared results
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
- Learning Less Is More - 6D Camera Localization via 3D Surface Regression (CVPR 2018 · similarity 0.88)
- Hierarchical Scene Coordinate Classification and Regression for Visual Localization (CVPR 2020 · similarity 0.87)
- Photo-SLAM: Real-time Simultaneous Localization and Photorealistic Mapping for Monocular Stereo and RGB-D Cameras (CVPR 2024 · similarity 0.86)
- Reloc3r: Large-Scale Training of Relative Camera Pose Regression for Generalizable, Fast, and Accurate Visual Localization (CVPR 2025 · similarity 0.86)
- VGGT: Visual Geometry Grounded Transformer (CVPR 2025 · similarity 0.86)
- SPAZER: Spatial-Semantic Progressive Reasoning Agent for Zero-shot 3D Visual Grounding (NeurIPS 2025 · similarity 0.85)
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.
