Research map / Multi-agent systems and games / Human-AI and goal-conditioned
CroCoDL: Cross-device Collaborative Dataset for 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.
Introduces CroCoDL, a multi-device visual localization dataset and benchmark with legged-robot (Spot), phone, HoloLens 2 and drone recordings across 10 LiDAR-aligned locations, evaluating cross-device localization with SfM-based pipelines.
Problem
Existing visual localization benchmarks focus on single-device scenarios, so cross-device localization between heterogeneous agents (robots, handheld phones, head-mounted MR headsets) with different viewpoints and sensors is under-evaluated.
Approach
Extends the LaMAR AR localization benchmark with Boston Dynamics Spot legged-robot recordings and new locations, generates pseudo-ground-truth by aligning all device sensor streams to NavVis LiDAR scans, and benchmarks cross-device map/query combinations with multiple retrieval, feature extraction and matching methods.
Key results
- Cross-device localization is harder than same-device localization across retrieval methods; iOS queries in Spot-built maps yield the lowest recall (e.g., 0.36 with APGeM + SuperPoint + LightGlue on Hydrology Lab).
- With ground-truth overlap retrieval (top 10), SuperPoint + LightGlue reaches 0.67 recall on the hardest Hydrology Lab scenario, showing retrieval is a major bottleneck but matching also struggles.
- The dataset comprises 10 locations, over 800 new sequences (about 1.5K total sequences), and more than 100 hours of raw recording time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace NetVLAD retrieval with SALAD | SuperPoint + LightGlue localization · Hydrology Lab, iOS query in Spot map | recall (5 deg, 0.5 m): 0.42 -> 0.54 | Helped |
| Use ground-truth overlap for top-10 retrieval instead of NetVLAD | SuperPoint + LightGlue localization · Hydrology Lab, iOS query in Spot map | recall (5 deg, 0.5 m): 0.42 -> 0.67 | Helped |
| Replace SuperPoint with SIFT | NetVLAD + LightGlue localization · Hydrology Lab, iOS query in Spot map | recall (5 deg, 0.5 m): 0.42 -> 0.25 | Hurt |
| Replace LightGlue with SuperGlue | NetVLAD + SuperPoint localization · Hydrology Lab, iOS query in Spot map | recall (5 deg, 0.5 m): 0.42 -> 0.43 | No clear effect |
| Replace LightGlue with LoFTR | NetVLAD + SuperPoint localization · Hydrology Lab, iOS query in Spot map | recall (5 deg, 0.5 m): 0.42 -> 0.40 | No clear effect |
| Use MASt3R instead of SuperPoint + LightGlue | Overlap top-1 retrieval localization · Hydrology Lab, iOS query in Spot map | recall (5 deg, 0.5 m): 0.57 -> 0.34 | Hurt |
Surprising findings
- Even with perfect (overlap-based) retrieval, feature matching for cross-device relative pose remains weak, e.g., MASt3R reaches only 0.34 recall in the hardest setting.
- GLACE, a coordinate regression method, achieves only a fraction of the NetVLAD-SuperPoint-LightGlue baseline on large-scale scenes.
Limitations
- Not all locations were fully covered by all devices; some narrow corridors and steep staircases were inaccessible to Spot.
- Drone data is not aligned to ground truth due to lack of exposed odometry.
- Spot registration into the NavVis scan is less accurate, so a larger recall threshold was used; iOS queries are limited to one image versus up to four for HoloLens 2 and Spot.
Experimental setup
Datasets: LaMAR CroCoDL
Benchmarks: CroCoDL
Metrics: recall at 5 degrees and 0.5 m recall at thresholds (0.5m to 5m, 10 deg)
Baselines: AP-GeM NetVLAD OpenIBL CosPlace SALAD LoFTR MASt3R GLACE
Builds on: LaMAR Hierarchical Localization LightGlue SuperPoint
Closest papers by meaning
- 360Loc: A Dataset and Benchmark for Omnidirectional Visual Localization with Cross-device Queries (CVPR 2024 · similarity 0.87)
- MAGiC-SLAM: Multi-Agent Gaussian Globally Consistent SLAM (CVPR 2025 · similarity 0.85)
- Seeing Across Skies and Streets: Feedforward 3D Reconstruction from Satellite, Drone, and Ground Images (NeurIPS 2026 · similarity 0.84)
- MCD: Diverse Large-Scale Multi-Campus Dataset for Robot Perception (CVPR 2024 · similarity 0.84)
- PIDLoc: Cross-View Pose Optimization Network Inspired by PID Controllers (CVPR 2025 · similarity 0.83)
- Enhancing Visual Localization with Cross-Domain Image Generation (ICML 2025 · similarity 0.83)
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.