Research map / 3D vision / Human pose estimation
PIDLoc: Cross-View Pose Optimization Network Inspired by PID Controllers
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 PIDLoc, a cross-view pose optimization network whose PID-inspired branches (P, I, D) and a spatially aware pose estimator refine vehicle pose from satellite images and LiDAR.
Problem
Cross-view vehicle pose optimization against satellite imagery degrades under large initial pose errors and repetitive patterns because existing methods only use cross-view features at the given pose.
Approach
Adds parallel PID-style branches: P uses feature differences at the current pose, I aggregates feature differences over a grid of pose candidates for global context, and D uses gradients of feature differences for fine-grained context; a channel-shared MLP spatial pose estimator (SPE) then jointly predicts the pose update from these features.
Key results
- On KITTI cross-area with ±20m/±10° initial error, PIDLoc reaches 4.96m mean position error vs 7.97m for Song et al. (about 37.8% reduction).
- On FMAVS cross-area, PIDLoc achieves 4.81m mean position error, versus 6.43m for PureACL (4C) and 15.20m for SIBCL.
- Ablation on KITTI cross-area shows the full PID set (PI+D) and SPE with positional embedding each improve lateral and longitudinal recall, e.g., SPE with PE reaches 71.01% lateral and 50.02% longitudinal recall@1m.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PIDLoc instead of Song et al. | cross-area KITTI pose optimization · KITTI cross-area | mean position error (m): 7.97 -> 4.96 | Helped |
| use PIDLoc instead of the strongest listed baseline | same-area FMAVS pose optimization · FMAVS same-area | mean position error (m): 4.19 -> 1.42 | Helped |
| use PIDLoc instead of PureACL (4C) | cross-area FMAVS pose optimization · FMAVS cross-area | mean position error (m): 6.43 -> 4.81 | Helped |
| add feature difference gradients in all directions to the PI branches (ablation) | PIDLoc D branch · KITTI cross-area | lateral recall R@0.25m (%): 19.03 -> 24.06 | Helped |
| add feature difference gradients in all directions to the PI branches (ablation) | PIDLoc D branch · KITTI cross-area | longitudinal recall R@0.25m (%): 13.54 -> 15.13 | Helped |
| add feature difference gradients in all directions to the PI branches (ablation) | PIDLoc D branch · KITTI cross-area | orientation recall R@0.25° (%): 58.34 -> 72.17 | Helped |
| add lateral gradients to the PI branches (ablation) | PIDLoc D branch · KITTI cross-area | lateral recall R@0.25m (%): 19.03 -> 22.67 | Helped |
| add longitudinal gradients to the PI branches (ablation) | PIDLoc D branch · KITTI cross-area | longitudinal recall R@0.25m (%): 13.54 -> 14.24 | Helped |
| add orientation gradients to the PI branches (ablation) | PIDLoc D branch · KITTI cross-area | orientation recall R@0.25° (%): 58.34 -> 69.52 | Helped |
| replace average-pooled features with SPE and positional embedding (ablation) | PIDLoc pose estimator · KITTI cross-area | longitudinal recall R@1m (%): 25.05 -> 50.02 | Helped |
| replace weighted-average pose estimation with SPE and positional embedding (ablation) | PIDLoc pose estimator · KITTI cross-area | longitudinal recall R@1m (%): 32.99 -> 50.02 | Helped |
| replace max-pooled features with SPE and positional embedding (ablation) | PIDLoc pose estimator · KITTI cross-area | orientation recall R@1° (%): 90.11 -> 99.96 | Helped |
Surprising findings
- The paper reports that prior methods' practice of aligning ground-truth poses to satellite image center may bias results, and reproduces baselines without this adjustment.
Limitations
- Relies on LiDAR points for sampling cross-view features and pose optimization.
- Initial pose error robustness is evaluated up to ±30m; the authors note future integration with SLAM to remove loop-closure dependency and improve initial pose accuracy.
Experimental setup
Datasets: KITTI Ford Multi-AV Seasonal (FMAVS)
Benchmarks: KITTI FMAVS
Metrics: mean position error (m) median position error (m) lateral recall@1m longitudinal recall@1m orientation error (deg) recall@1 deg
Baselines: DSM HighlyAccurate SliceMatch PureACL VFA SIBCL Boosting CCVPE Song et al.
Builds on: SIBCL HighlyAccurate PureACL VFA Boosting
Closest papers by meaning
- View From Above: Orthogonal-View aware Cross-view Localization (CVPR 2024 · similarity 0.85)
- Seeing Across Skies and Streets: Feedforward 3D Reconstruction from Satellite, Drone, and Ground Images (NeurIPS 2026 · similarity 0.85)
- BevSplat: Resolving Height Ambiguity via Feature-Based Gaussian Primitives for Weakly-Supervised Cross-View Localization (NeurIPS 2025 · similarity 0.84)
- Weather-Robust Cross-View Geo-Localization via Prototype-Based Semantic Part Discovery (NeurIPS 2026 · similarity 0.84)
- LoD-Loc: Aerial Visual Localization using LoD 3D Map with Neural Wireframe Alignment (NeurIPS 2024 · similarity 0.84)
- Aligning Geometric Spatial Layout in Cross-View Geo-Localization via Feature Recombination (AAAI 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.