Research map / Vision-language and multimodal / Human and embodied modeling
Game4Loc: A UAV Geo-Localization Benchmark from Game Data
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.
Builds GTA-UAV, a 33,763-image drone-view dataset from GTA V with partial (IoU-based) drone-satellite matches, and trains a ViT dual encoder with a weighted-InfoNCE loss plus mutually exclusive sampling for UAV geo-localization.
Problem
UAV geo-localization datasets assume perfect one-to-one aligned drone-satellite pairs, which does not match practical scenarios where partial overlaps are common; existing real-world data is small and limited in scenes and altitudes.
Approach
Pair drone and satellite tiles by ground-coverage IoU (positive >0.39, semi-positive 0.14-0.39) rather than center alignment, and weight the InfoNCE positive term by a sigmoid of the IoU so partial matches are pulled together with graded strength; a mutually exclusive batch sampler avoids related pairs in the same batch.
Key results
- On GTA-UAV cross-area with positive+semi-positive training, weighted-InfoNCE reaches 55.91% R@1 and 342.05m Dis@1, versus 45.97% R@1 and 460.08m for InfoNCE with mutually exclusive sampling.
- Models pre-trained on GTA-UAV and fine-tuned on UAV-VisLoc reach 80.20% R@1 and 122.87m Dis@1, outperforming ImageNet, University-1652, SUES-200 and DenseUAV pre-training.
- In zero-shot cross-area transfer to UAV-VisLoc, GTA-UAV pre-training gives 24.94% R@1 versus at most 18.79% for the other pre-training datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace InfoNCE with weighted-InfoNCE while keeping mutually exclusive sampling and positive plus semi-positive training pairs | GTA-UAV training with ViT-Base/16 · GTA-UAV cross-area | R@1: 45.97% -> 55.91% | Helped |
| Replace InfoNCE with weighted-InfoNCE while keeping mutually exclusive sampling and positive plus semi-positive training pairs | GTA-UAV training with ViT-Base/16 · GTA-UAV cross-area | Dis@1: 460.08m -> 342.05m | Helped |
| Replace InfoNCE with weighted-InfoNCE while keeping mutually exclusive sampling and positive plus semi-positive training pairs | GTA-UAV training with ViT-Base/16 · GTA-UAV same-area | R@1: 65.89% -> 84.95% | Helped |
| Replace InfoNCE with weighted-InfoNCE while keeping mutually exclusive sampling and positive plus semi-positive training pairs | GTA-UAV training with ViT-Base/16 · GTA-UAV same-area | Dis@1: 196.59m -> 149.07m | Helped |
| Train with positive and semi-positive pairs instead of positive-only pairs | Weighted-InfoNCE with MES on GTA-UAV · GTA-UAV cross-area | R@1: 57.52% -> 55.91% | Mixed |
| Train with positive and semi-positive pairs instead of positive-only pairs | Weighted-InfoNCE with MES on GTA-UAV · GTA-UAV cross-area | Dis@1: 444.13m -> 342.05m | Helped |
| Add mutually exclusive sampling to InfoNCE training | GTA-UAV with positive plus semi-positive pairs · GTA-UAV cross-area | R@1: 35.83% -> 45.97% | Helped |
| Pre-train on GTA-UAV rather than SUES-200 | ViT-Base/16 evaluated on GTA-UAV · GTA-UAV cross-area | R@1: 35.29% -> 55.91% | Helped |
| Pre-train on GTA-UAV rather than DenseUAV | ViT-Base/16 transferred to UAV-VisLoc without fine-tuning · UAV-VisLoc same-area, zero-shot | R@1: 18.79% -> 24.94% | Helped |
| Pre-train on GTA-UAV rather than DenseUAV before fine-tuning | ViT-Base/16 transferred to UAV-VisLoc · UAV-VisLoc same-area, fine-tune | R@1: 77.09% -> 80.20% | Helped |
| Use ViT-Base/16 rather than ConvNeXt-Base (ablation) | GTA-UAV localization · GTA-UAV cross-area | R@1: 55.36% -> 55.91% | Helped |
| Set weighted-InfoNCE k to 5 rather than 1, 20, or infinity (ablation) | Weighted-InfoNCE on GTA-UAV · GTA-UAV cross-area | R@1: 52.96% -> 55.91%; 51.50% -> 55.91%; 45.97% -> 55.91% | Helped |
Surprising findings
- Training on positive plus semi-positive pairs lowers R@1 relative to positive-only training, though it improves localization metrics SDM@3 and Dis@1.
- ResNet-101 generalizes poorly in cross-area settings (13.74% R@1) compared with ViT-Base/16 (55.91%).
Limitations
- Real-world evaluation relies on a single real dataset (UAV-VisLoc) with limited scale and scenes.
- Dataset is synthetic from a commercial game, so domain gap to real imagery is a concern.
Experimental setup
Datasets: GTA-UAV University-1652 SUES-200 DenseUAV UAV-VisLoc
Benchmarks: GTA-UAV UAV-VisLoc
Metrics: Recall@K AP SDM@3 Dis@1 (meters)
Baselines: Triplet loss InfoNCE loss Sample4Geo ImageNet pre-training University-1652 pre-training SUES-200 pre-training DenseUAV pre-training ResNet-101 ConvNeXt-Base Swinv2-B
Builds on: InfoNCE Sample4Geo VIGOR
Closest papers by meaning
- Weather-Robust Cross-View Geo-Localization via Prototype-Based Semantic Part Discovery (NeurIPS 2026 · similarity 0.91)
- OrthoLoC: UAV 6-DoF Localization and Calibration Using Orthographic Geodata (NeurIPS 2025 · similarity 0.88)
- SatNav: A Scalable Benchmark for Long-Horizon UAV Vision-Language Navigation from Satellite Imagery (NeurIPS 2026 · similarity 0.87)
- Seeing Across Skies and Streets: Feedforward 3D Reconstruction from Satellite, Drone, and Ground Images (NeurIPS 2026 · similarity 0.86)
- From Coarse to Fine: A Matching and Alignment Framework for Unsupervised Cross-View Geo-Localization (AAAI 2025 · similarity 0.86)
- Towards Realistic UAV Vision-Language Navigation: Platform, Benchmark, and Methodology (ICLR 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.