Research map / 3D vision / Depth and stereo
GLACE: Global Local Accelerated Coordinate Encoding
Read the paperCheck your idea against 108,967 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.
Scales scene coordinate regression to large scenes with a single small network by feature-diffusing pretrained global retrieval encodings and using a clustered position decoder, without 3D supervision.
Problem
Scene coordinate regression (SCR) works well in small scenes but struggles at large scale when trained only with reprojection loss (no 3D ground truth), because of ambiguity between similar-looking unrelated views and implicit triangulation difficulty.
Approach
Concatenate pretrained global image-retrieval features (R2Former) with ACE local encodings, perturb the global features with Gaussian noise (feature diffusion) so the network groups reprojection constraints by co-visibility, and replace ACE's unimodal center-based output with a softmax convex combination over K-Means cluster centers of camera positions.
Key results
- On Cambridge Landmarks, GLACE with a single 13MB model achieves 14 cm / 0.3 deg average median error, versus 17 cm for Poker (ACE x4, 16MB).
- On Aachen Day, GLACE (9MB map noted in paper, 27MB in Table 5) reaches 20.8% at 0.5m/5deg versus 17.2% for ACE x50 (205MB).
- On i19 integrated rooms with SfM poses, GLACE (9MB) reaches 93.4% at 5cm/5deg versus 97.8% for ACE x19 (78MB) and 42.0% for ACE x4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GLACE instead of Poker (ACE × 4) | SCR on Cambridge Landmarks · Cambridge Landmarks | position error (cm): 17 -> 14 | Helped |
| Use GLACE instead of ACE | SCR on Cambridge Landmarks · Cambridge Landmarks | position error (cm): 25 -> 14 | Helped |
| Use GLACE instead of ACE × 50 | SCR on Aachen Day · Aachen Day | accuracy at 0.25m,2°: 6.9% -> 8.6% | Helped |
| Use GLACE instead of ACE × 19 | SCR on integrated rooms i19 with SfM poses · Integrated rooms i19, SfM poses | accuracy below 5cm, 5°: 97.8% -> 93.4% | Hurt |
| Add feature diffusion rather than identity global encoding (ablation) | GLACE on Cambridge Landmarks · Cambridge Landmarks | position error (cm): 0.8 -> 0.3 | Helped |
| Replace feature diffusion with K-Means global encoding (K=4) (ablation) | GLACE on Cambridge Landmarks · Cambridge Landmarks | position error (cm): 0.3 -> 0.3 | No clear effect |
| Replace feature diffusion with K-Means global encoding (K=32) (ablation) | GLACE on Cambridge Landmarks · Cambridge Landmarks | position error (cm): 0.3 -> 0.3 | No clear effect |
| Replace feature diffusion with K-Means global encoding (K=128) (ablation) | GLACE on Cambridge Landmarks · Cambridge Landmarks | position error (cm): 0.3 -> 0.3 | No clear effect |
| Increase decoder cluster count from K=1 to K=4 (ablation) | GLACE on integrated rooms i19 · Integrated rooms i19, SfM poses | accuracy below 5cm, 5°: 57.5% -> 82.6% | Helped |
| Increase decoder cluster count from K=4 to K=19 (ablation) | GLACE on integrated rooms i19 · Integrated rooms i19, SfM poses | accuracy below 5cm, 5°: 82.6% -> 89.3% | Helped |
| Increase decoder cluster count from K=19 to K=50 (ablation) | GLACE on integrated rooms i19 · Integrated rooms i19, SfM poses | accuracy below 5cm, 5°: 89.3% -> 93.4% | Helped |
Surprising findings
- Naively concatenating global encodings to local encodings causes overfitting, notably on simple scenes like StMarysChurch (103 cm error vs 9 cm with diffusion).
Limitations
- Large-scene performance still trails ensemble ACE x19 on i19 with SfM poses (93.4% vs 97.8%).
- Performance on GreatCourt remains only comparable to feature-matching methods rather than exceeding them.
- Requires distributed multi-GPU training with large batch sizes and buffers to cache training data.
Experimental setup
Datasets: Cambridge Landmarks 7 Scenes 12 Scenes Aachen Day-Night MSLS
Benchmarks: 7 Scenes 12 Scenes Integrated Rooms (i12, i19) Cambridge Landmarks Aachen Day
Metrics: percentage of frames below 5cm/5deg median rotation and position error accuracy at 0.25m/2deg, 0.5m/5deg, 5m/10deg
Baselines: ACE Poker (ACE x4) DSAC* ESAC hLoc (SP+SG) AS (SIFT) pixLoc GoMatch SANet SRC
Builds on: ACE R2Former DSAC* hLoc
Closest papers by meaning
- R-SCoRe: Revisiting Scene Coordinate Regression for Robust Large-Scale Visual Localization (CVPR 2025 · similarity 0.92)
- GS-CPR: Efficient Camera Pose Refinement via 3D Gaussian Splatting (ICLR 2025 · similarity 0.86)
- Scene-agnostic Pose Regression for Visual Localization (CVPR 2025 · similarity 0.85)
- Reloc3r: Large-Scale Training of Relative Camera Pose Regression for Generalizable, Fast, and Accurate Visual Localization (CVPR 2025 · similarity 0.85)
- Map-Relative Pose Regression for Visual Re-Localization (CVPR 2024 · similarity 0.85)
- Learning to Produce Semi-dense Correspondences for Visual Localization (CVPR 2024 · 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.