SCC-Loc: A Unified Semantic Cascade Consensus Framework for UAV Thermal Geo-Localization
| Authors | Xiaoran Zhang et al. |
| Year | 2026 |
| Field | Computer Vision |
| arXiv | 2604.03120 |
| Download | |
| Categories | cs.CV, cs.RO |
Abstract
Cross-modal Thermal Geo-localization (TG) provides a robust, all-weather solution for Unmanned Aerial Vehicles (UAVs) in Global Navigation Satellite System (GNSS)-denied environments. However, profound thermal-visible modality gaps introduce severe feature ambiguity, systematically corrupting conventional coarse-to-fine registration. To dismantle this bottleneck, we propose SCC-Loc, a unified Semantic-Cascade-Consensus localization framework. By sharing a single DINOv2 backbone across global retrieval and MINIMA_{\text{RoMa}} matching, it minimizes memory footprint and achieves zero-shot, highly accurate absolute position estimation. Specifically, we tackle modality ambiguity by introducing three cohesive components. First, we design the Semantic-Guided Viewport Alignment (SGVA) module to adaptively optimize satellite crop regions, effectively correcting initial spatial deviations. Second, we develop the Cascaded Spatial-Adaptive Texture-Structure Filtering (C-SATSF) mechanism to explicitly enforce geometric consistency, thereby eradicating dense cross-modal outliers. Finally, we propose the Consensus-Driven Reliability-Aware Position Selection (CD-RAPS) strategy to derive the optimal solution through a synergy of physically constrained pose optimization. To address data scarcity, we construct Thermal-UAV, a comprehensive dataset providing 11,890 diverse thermal queries referenced against a large-scale satellite ortho-photo and corresponding spatially aligned Digital Surface Model (DSM). Extensive experiments demonstrate that SCC-Loc establishes a new state-of-the-art, suppressing the mean localization error to 9.37 m and providing a 7.6-fold accuracy improvement within a strict 5-m threshold over the strongest baseline. Code and dataset are available at https://github.com/FloralHercules/SCC-Loc.
Engineering Breakdown
Plain English
This paper addresses thermal-visible image geo-localization for UAVs operating without GPS in all-weather conditions. The core problem is that thermal and visible spectrum images have fundamentally different appearances, making it hard to match them for localization. The authors propose SCC-Loc, which uses a shared DINOv2 backbone to bridge this modality gap through semantic guidance and cascade consensus matching, enabling zero-shot accurate absolute position estimation without requiring paired thermal-visible training data.
Core Technical Contribution
The key innovation is a unified framework that tackles cross-modal feature ambiguity through three integrated components: semantic-guided viewport alignment, cascade consensus matching (MINIMA-RoMa), and a shared DINOv2 backbone that eliminates redundant computation. Unlike traditional coarse-to-fine registration pipelines that fail on thermal-visible pairs due to appearance mismatch, SCC-Loc leverages semantic features from vision transformers that are inherently more modality-agnostic. The framework achieves this with zero-shot capability, meaning it requires no thermal-visible paired training data, which is significant for practical deployment where such datasets are expensive to collect.
How It Works
The pipeline starts with a thermal query image and a database of georeferenced visible images. First, semantic-guided viewport alignment uses shared DINOv2 features to retrieve candidate visible images from global regions without pixel-level matching—this is the coarse retrieval step that filters the massive database down to a manageable set. Second, the MINIMA-RoMa matching component performs fine-grained correspondence between the thermal query and retrieved visible candidates by enforcing geometric consistency and consensus constraints. Third, the cascade consensus mechanism validates correspondences across multiple levels (semantic, geometric, photometric) to reject spurious matches and enforce agreement. The final output is an absolute geographic coordinate derived from the validated correspondences, with confidence scores quantifying localization certainty.
Production Impact
For UAV systems operating in GPS-denied environments (underground, urban canyons, jamming scenarios), this approach provides a practical alternative to inertial navigation and visual odometry, which accumulate drift over time. The single shared DINOv2 backbone reduces memory footprint compared to dual-network architectures, making deployment feasible on edge devices with 2-4GB GPU memory—critical for battery-constrained aerial platforms. Integration requires pre-computed DINOv2 features for reference visible imagery (fast offline step) and real-time feature extraction for thermal queries (typically 50-200ms on modern GPUs). The zero-shot capability eliminates the need to fine-tune on mission-specific thermal-visible pairs, reducing deployment overhead. Trade-offs include dependency on visual distinctiveness (fails in repetitive terrain like deserts) and requirement for properly georeferenced reference imagery database.
Limitations and When Not to Use This
The approach assumes reference imagery is available and accurately georeferenced, which may not hold in rapidly changing environments or when maps are outdated. Performance degrades significantly in textureless or repetitive scenes where semantic features provide weak discriminative information, as the method fundamentally relies on visual distinctiveness. The paper doesn't thoroughly evaluate failure modes under extreme thermal conditions (intense heat sources, frost) or at scale across continental-sized regions where reference imagery sparsity becomes critical. Computational bottlenecks during candidate retrieval (for multi-million image databases) and runtime latency guarantees for real-time UAV navigation are not quantified. The approach also lacks comparison against recent learning-based localization methods that explicitly train on thermal-visible pairs, so relative performance gains remain unclear.
Research Context
This work builds on the foundation of DINOv2 for zero-shot visual understanding and recent advances in cross-modal matching (RoMa, MINIMA variants). It advances the specific problem of thermal-visible registration, which is relatively underexplored compared to visible-infrared matching in pedestrian detection. The contribution fits within broader research on GNSS-denied localization for autonomous systems, complementing simultaneous localization and mapping (SLAM) and neural radiance field (NeRF) approaches. It opens research directions in cross-spectral semantic alignment and demonstrates that foundation models trained on visible imagery can transfer effectively to thermal domains without fine-tuning—relevant for multimodal perception in robotics and autonomous vehicles.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::