Research map / 3D vision / Depth and stereo

Weather-Robust Cross-View Geo-Localization via Prototype-Based Semantic Part Discovery

NeurIPS 2026 · 3D vision › Depth and stereo · approach: Scene coordinate regression for localization · 0 citations (OpenAlex)

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 SkyPart, a lightweight prototype-based part-discovery head on DINOv2 ViT-S/14 for drone-to-satellite cross-view geo-localization, with altitude FiLM used only in training and Kendall-weighted multi-task loss, reporting SOTA on SUES-200, University-1652, and DenseUAV under a single-pass protocol.

Problem

Cross-view geo-localization of drones against satellite tiles is hurt by global descriptors that discard layout structure, altitude-induced scale variation, and hand-tuned multi-objective loss weighting, and performance degrades under weather corruption.

Approach

A swappable head for patch-based ViTs that groups patch tokens into learnable prototypes via single-pass cosine competition, fuses CLS, part, and graph (GAT over prototypes) readouts, applies altitude FiLM only during training with mean-FiLM at inference, and uses Kendall uncertainty weighting over four loss groups.

Key results

• On SUES-200 D->S, SkyPart reaches 99.60 mean R@1 at 26.95M params and 22.14 GFLOPs, ranking first across altitudes.
• On University-1652 D->S, SkyPart achieves 96.47 R@1 and 97.77 AP, versus 94.46 R@1 for CAMP.
• Under WeatherPrompt corruptions on SUES-200, SkyPart reports 94.71 mean D->S R@1 versus 62.52 for WeatherPrompt (baselines reported at 384px vs SkyPart at 392px).

Surprising findings

• Removing the alignment loss group collapses Mean R@1 by about 68 points in the ablation, while removing any auxiliary group costs only about one point.
• Forward-KL variant of UAPA underperforms reverse KL by 2.22 pp R@1.

Limitations

• Baseline numbers are taken from original papers under differing protocols and resolutions, so comparisons are not fully controlled.
• WeatherPrompt SUES-200 and University-1652 baselines use 384px while SkyPart is evaluated at 392px.
• Ablation checkpoints differ from the deployed checkpoint, and the paper notes the Pareto-stationarity guarantee does not imply Pareto-optimality.

Experimental setup

Datasets: SUES-200 University-1652 DenseUAV WeatherPrompt corruptions

Benchmarks: SUES-200 University-1652 DenseUAV WeatherPrompt

Metrics: R@1 AP R@5 SDM@1 mean R@1 under corruption

Baselines: LPN Sample4Geo MCCG SeGCN CCR CAMP DAC MEAN WeatherPrompt Safe-Net DenseUAV baseline CEUSP

Builds on: DINOv2 slot attention Kendall homoscedastic uncertainty weighting Graph Attention Networks Sample4Geo

Closest papers by meaning

• Game4Loc: A UAV Geo-Localization Benchmark from Game Data (AAAI 2025 · similarity 0.91)
• UniGeoRS: A Unified Benchmark for Tri-view Geo-Localization (CVPR 2026 · similarity 0.90)
• Seeing Across Skies and Streets: Feedforward 3D Reconstruction from Satellite, Drone, and Ground Images (NeurIPS 2026 · similarity 0.89)
• WeatherPrompt: Multi-modality Representation Learning for All-Weather Drone Visual Geo-Localization (NeurIPS 2025 · similarity 0.87)
• From Coarse to Fine: A Matching and Alignment Framework for Unsupervised Cross-View Geo-Localization (AAAI 2025 · similarity 0.86)
• BevSplat: Resolving Height Ambiguity via Feature-Based Gaussian Primitives for Weakly-Supervised Cross-View Localization (NeurIPS 2025 · similarity 0.86)