Research map / 3D vision / Depth and stereo
F3Loc: Fusion and Filtering for Floorplan Localization
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.
Proposes F3Loc, a data-driven floorplan localization system that fuses monocular and multi-view floorplan depth predictions with a learned selection network and an SE(2) histogram filter for sequential localization.
Problem
Self-localization of a camera from RGB images within a 2D floorplan, handling ambiguity from repetitive structures, non-upright camera poses, and real-time sequential tracking without per-map retraining or image databases.
Approach
Predicts 1D floorplan depth rays from single images (gravity-aligned) and from multi-view cost volumes, fuses them with an MLP selection network conditioned on relative poses, applies virtual roll-pitch augmentation for non-upright cameras, and filters over time with an efficient SE(2) histogram filter implemented as grouped convolution.
Key results
- Monocular module on Gibson(f) reaches 36.6% recall at 1m vs 13.0% for LASER and 1.2% for PF-net (Table 1).
- Full system with selection network reaches 94.6% sequential success rate at 1m vs 59.5% for LASER and 89.2% for monocular-only filtering (Table 3).
- Selection network improves recall by 30-50% over individual monocular and multi-view networks on Gibson(g) (Table 2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use multiview depth estimation instead of monocular estimation (ablation) | F3Loc observation model · Gibson(f) | recall@1m: 36.6 -> 45.2 | Helped |
| use multiview depth estimation instead of monocular estimation (ablation) | F3Loc observation model · Gibson(g) | recall@1m: 33.7 -> 31.0 | Hurt |
| add learned selection between monocular and multiview estimates (ablation) | F3Loc observation model · Gibson(g) | recall@1m: 33.7 -> 44.5 | Helped |
| select estimates by thresholding relative motion rather than learned selection (ablation) | F3Loc observation model · Gibson(g) | recall@1m: 39.6 -> 44.5 | Hurt |
| add virtual roll-pitch augmentation during monocular network training (ablation) | F3Loc monocular module · Gibson(f), non-upright camera poses | recall@1m 30° | Helped |
| increase the number of frames used by the histogram filter (ablation) | F3Loc sequential localization · Gibson(t) | success rate | Helped |
| replace learned selection with monocular observation (ablation) | F3Loc histogram-filtered localization · Gibson(t) | success rate@1m: 94.6 -> 89.2 | Hurt |
| use the histogram filter instead of the particle filter (ablation) | F3Loc sequential localization · Gibson(t) | iteration time: 0.067 -> 0.037 | Helped |
Surprising findings
- Multi-view estimation recall falls below monocular on Gibson(g), which includes near in-place rotation with small baselines.
Limitations
- Relies on simulated iGibson data; domain gap to real-world imagery remains.
- Floorplan ambiguities (repetitive structures like hallways) remain without semantic cues.
- Multi-view estimation degrades with small baselines or insufficient overlap.
Experimental setup
Datasets: iGibson (customized Gibson-based floorplan dataset) Structured3D LaMAR (HGE scene)
Benchmarks: Gibson(f) Gibson(g) Gibson(t) Structured3D LaMAR HGE
Metrics: recall at 0.1m/0.5m/1m/1m30deg success rate@1m RMSE runtime (s, Hz)
Baselines: PF-net LASER LaLaLoc
Builds on: PF-net LASER histogram filter (Bayesian filter with transition filters) MVSNet-style multi-view stereo ResNet
Closest papers by meaning
- UnLoc: Leveraging Depth Uncertainties for Floorplan Localization (ICLR 2026 · similarity 0.92)
- Fusion of Depth and Semantics for Probabilistic Floorplan Localization (CVPR 2026 · similarity 0.90)
- CAGE: Continuity-Aware edGE Network Unlocks Robust Floorplan Reconstruction (NeurIPS 2025 · similarity 0.83)
- HouseLayout3D: A Benchmark and Training-free Baseline for 3D Layout Estimation in the Wild (NeurIPS 2025 · similarity 0.83)
- 360Loc: A Dataset and Benchmark for Omnidirectional Visual Localization with Cross-device Queries (CVPR 2024 · similarity 0.82)
- Rooms from Motion: Un-posed Indoor 3D Object Detection as Localization and Mapping (NeurIPS 2025 · similarity 0.82)
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.