---
title: 'GLEAM-C: Unified Cross-View Geo-Localization'
url: https://www.emergentmind.com/topics/gleam-c
type: topic
---

# GLEAM-C: Unified Cross-View Geo-Localization

GLEAM-C is the “core” part of the GLEAM framework: a foundational cross-view geo-localization model for Cross-View Geo-Localization (CVGL) that learns a single visual embedding space in which diverse views and modalities of the same place are close, and mismatched locations are far apart. Its task is to take a query image captured from one viewpoint or modality and retrieve the corresponding satellite image from a geo-referenced aerial database, with the best match providing the estimated geographic position. Within the broader GLEAM framework, GLEAM-C handles matching, while GLEAM-X introduces explainable reasoning on top of the matching process [2509.07450].

## 1. Problem formulation and satellite-centric alignment

GLEAM-C addresses CVGL in a multi-view, multi-modality setting rather than the more usual single query view versus satellite formulation. The input is a query image from one of several views or modalities, the reference database consists of tiled and geo-referenced satellite imagery, and the output is a ranked list of satellite images or a top-1 prediction. This preserves the standard retrieval formulation of CVGL while extending it to a single unified model that can process multiple query types [2509.07450].

A defining design choice is that all non-satellite modalities are aligned only to satellite imagery. There is no direct learning of ground-to-UAV or ground-to-map correspondences. Instead, every modality is trained to map into a shared satellite-anchored embedding. This produces a single embedding space in which UAV imagery, panoramic ground views, ground-level perspective photographs, and street maps are all aligned through their corresponding satellite images.

This satellite-centric organization has two immediate consequences. First, it permits a unified training recipe across heterogeneous datasets without modality-specific pairwise objectives. Second, it makes satellite imagery the reference hub for retrieval and deployment. This suggests that GLEAM-C is best understood not as a collection of pairwise CVGL models, but as a single retrieval backbone whose geometry is defined by satellite correspondences.

## 2. Supported modalities and dataset composition

GLEAM-C is trained on four major non-satellite modalities, all paired with satellite images. The supported query types are UAV imagery, panoramic ground views, standard ground-level perspective photographs, and street maps. Satellite imagery functions as the central domain in all cases [2509.07450].

| Non-satellite modality | Dataset | Retrieval task |
|---|---|---|
| UAV imagery | University-1652 | drone-image to satellite |
| Panoramic ground views | VIGOR | panorama to satellite |
| Ground-level perspective photos | SetVL-480K | ground photo to satellite |
| Street maps | MAP | map tile to satellite |

The UAV branch uses University-1652 under the Drone2Sat protocol, with oblique or near-nadir aerial images from UAVs. Panoramic ground views are drawn from VIGOR, which contains 360° panoramas captured at ground level. Standard perspective ground photographs are taken from SetVL-480K, identified in the data as originating from Wu et al., “Cross-View Image Set Geo-Localization.” The map modality is based on the MAP dataset collected by the authors, consisting of 12,761 pairs of street map tiles and matching satellite images from Google Maps, split into 10,208 for training and 2,553 for test [2509.07450].

The training composition is explicitly balanced by dataset-level sampling ratios because the original datasets are highly imbalanced in size. The reported sampled counts are 37,854 pairs for University-1652 with sampling ratio 1.0, 52,609 for VIGOR with sampling ratio 1.0, 120,272 for SetVL-480K with sampling ratio 0.5, and 40,832 for MAP with sampling ratio 4.0. The total number of training pairs used for GLEAM-C is 251,567. The oversampling of MAP and downsampling of SetVL-480K are intended to ensure that each modality contributes meaningfully during training rather than allowing the largest dataset to dominate.

## 3. Model architecture and embedding objective

Architecturally, GLEAM-C is deliberately simple and is described as essentially a CLIP-style image-image contrastive model. It uses one image encoder for all query views and one for satellite reference images in a symmetric contrastive setup; in practice, the two branches share the same architecture and weights, as in CLIP-style contrastive learning. There is no modality-specific encoder, no cross-attention, no fusion module, and no additional modality-specific alignment term [2509.07450].

Two image encoder backbones are reported. One is ConvNeXt-B-384, a CNN with 88.6M parameters. The other is PE-Core-L14-336, described as a “Perception Encoder” ViT-like model with approximately 0.32B parameters. For each configuration, the same backbone is used for all modalities and for both query and reference images.

Given an input image $x$, the encoder $f_\theta$ produces a $d$-dimensional embedding
$$
\mathbf{f} = f_\theta(x) \in \mathbb{R}^d.
$$
For a batch of query images and their corresponding satellite images, the features are L2-normalized so that cosine similarity reduces to a dot product:
$$
\mathbf{f}_1 \leftarrow \text{normalize}(\mathbf{f}_1), \quad \mathbf{f}_2 \leftarrow \text{normalize}(\mathbf{f}_2).
$$

The loss is a symmetric InfoNCE or CLIP-style contrastive objective. Features are gathered across all GPUs, yielding
$$
\mathbf{F}_1 \in \mathbb{R}^{N \times d}, \quad \mathbf{F}_2 \in \mathbb{R}^{N \times d},
$$
with $N = B \times W$ for batch size $B$ and world size $W$. Logits are computed as
$$
\mathbf{Z}_1 = \alpha \cdot \mathbf{F}_1 \mathbf{F}_2^\top, \quad \mathbf{Z}_2 = \mathbf{Z}_1^\top,
$$
where $\alpha$ is an optional learnable logit scale. Labels are
$$
\mathbf{y} = [0, 1, \dots, N-1],
$$
and the final symmetric loss is
$$
\mathcal{L} = \frac{1}{2} \big( \text{CE}(\mathbf{Z}_1, \mathbf{y}) + \text{CE}(\mathbf{Z}_2, \mathbf{y}) \big).
$$

The paper also gives the equivalent pairwise form
$$
\mathbf{Z}_{ij} = \alpha \cdot \langle \mathbf{f}_1^{(i)}, \mathbf{f}_2^{(j)} \rangle,
$$
with
$$
\mathcal{L}_{1} = - \frac{1}{N} \sum_{i=1}^N \log \frac{\exp(\mathbf{Z}_{ii})}{\sum_{j=1}^N \exp(\mathbf{Z}_{ij})},
$$
$$
\mathcal{L}_{2} = - \frac{1}{N} \sum_{i=1}^N \log \frac{\exp(\mathbf{Z}_{ii})}{\sum_{j=1}^N \exp(\mathbf{Z}_{ji})},
$$
and
$$
\mathcal{L} = \frac{1}{2}(\mathcal{L}_{1} + \mathcal{L}_{2}).
$$

All cross-view alignment is therefore achieved implicitly through shared-encoder contrastive learning on query-satellite pairs. A plausible implication is that GLEAM-C treats modality diversity primarily as a data problem rather than an architectural one.

## 4. Training curriculum, sampling, and optimization

The paper identifies naïve from-scratch multi-dataset training as unstable and imbalanced. In particular, the encoder initially has weak matching capacity, and the varying difficulty and size of the datasets lead to performance degradation on some views, especially VIGOR. To address this, GLEAM-C adopts a two-phase curriculum [2509.07450].

Phase 1 is single-dataset pretraining on VIGOR for 40 epochs. VIGOR is characterized as relatively large and of moderate difficulty, providing good coverage of ground-satellite variation. Phase 2 initializes from the VIGOR checkpoint and fine-tunes on the merged sampled datasets—VIGOR, University-1652, SetVL-480K, and MAP—for another 40 epochs, using the same symmetric InfoNCE objective.

Three strategies are compared: single-dataset training, from-scratch Merge, and the two-phase VIGOR-to-Merge curriculum. The two-phase strategy is the adopted configuration because it systematically outperforms from-scratch Merge and yields performance comparable to or better than single-dataset baselines while retaining a unified model.

GLEAM-C is built on the Sample4Geo framework and inherits Dynamic Similarity Sampling (DSS). Sample4Geo includes GPS-based sampling and DSS, but because some datasets such as University-1652 lack GPS metadata, GLEAM-C cannot use GPS-based sampling and therefore relies on DSS only. Within each batch, DSS chooses dynamic hard negatives based on similarity distributions. This is used in place of random negative sampling to improve training efficiency.

For PE-Core-L14-336 in Phase 2, the reported hyperparameters are 40 epochs, total batch size 300 across GPUs, learning rate $1 \times 10^{-4}$, cosine scheduler, warmup epochs 1, label smoothing 0.1, neighbor select 64, neighbor range 128, rotation probability 0.75, and flip probability 0.5. The paper characterizes these as CLIP-style settings tailored for contrastive image-image training.

## 5. Retrieval procedure and empirical performance

At inference time, GLEAM-C functions as a standard retrieval-based CVGL model. All satellite tiles in the reference database are encoded with $f_\theta$ to obtain normalized satellite embeddings, which may be stored in a vector index. A query image from any supported modality is then encoded to a normalized embedding, cosine similarity is computed against the satellite database, and the resulting ranking is used for geo-localization. The paper notes that an approximate nearest neighbor index such as FAISS is standard usage, but does not specify an ANN engine [2509.07450].

Evaluation follows the conventions of the underlying datasets. VIGOR uses Hit Rate and Recall@1/5/10 and considers candidate regions in a local area around ground truth. University-1652 under Drone2Sat reports Recall@1/5/10 and Average Precision. SetVL-480K uses set-level retrieval metrics such as Recall@K and Hit Rate. MAP uses evaluation logic similar to VIGOR.

The central empirical result is that a single unified model can approach or exceed the performance of modality-specific baselines. With the ConvNeXt-B-384 backbone, two-phase training yields VIGOR $R@1 = 75.66$ and Hit Rate $= 86.80$, University-1652 $R@1 = 87.03$ and $AP = 89.12$, MAP $R@1 = 94.05$ and $HR = 94.05$, and SetVL-480K $R@1 = 15.28$ and $HR = 15.28$. Relative to from-scratch Merge, this recovers much of the VIGOR degradation and improves MAP and SetVL-480K. Relative to single-dataset training, MAP and SetVL-480K improve while VIGOR remains close [2509.07450].

With the larger PE-Core-L14-336 backbone, two-phase training produces VIGOR $R@1 = 75.96$ and $HR = 89.44$, University-1652 $R@1 = 93.19$ and $AP = 94.28$, MAP $R@1 = 93.97$ and $HR = 93.97$, and SetVL-480K $R@1 = 23.25$ and $HR = 23.25$. The larger PE-Core model outperforms ConvNeXt-B across all datasets, and the two-phase curriculum is described as crucial for avoiding degradation on VIGOR when datasets are merged.

The paper summarizes these results as near-SOTA performance comparable to modality-specific models, with minor differences attributed to the absence of GPS-based sampling. This suggests that unification in GLEAM-C is not primarily achieved by sacrificing task-specific quality, but by organizing training around a shared satellite reference space.

## 6. Implementation efficiency, relation to GLEAM-X, and limitations

The reported efficiency improvements are training-system changes rather than architectural changes. Sample4Geo originally uses PyTorch Data Parallel, which centralizes communication and creates a bottleneck on GPU 0. GLEAM-C replaces this with Distributed Data Parallel and OpenCLIP-style global negative gathering: each GPU computes gradients locally, then all feature vectors are AllGathered before the InfoNCE loss is computed so that each device sees the global batch of negatives [2509.07450].

The empirical speedup reported on VIGOR with ConvNeXt-B is from 521 seconds per epoch under Data Parallel to 92 seconds per epoch under Distributed Data Parallel, described as more than 5× faster with essentially the same accuracy. The paper also identifies the use of a single encoder across modalities and a unified contrastive head as deployment advantages because they reduce memory footprint and avoid modality-specific branches or fusion modules.

GLEAM-C is one half of the full GLEAM framework. GLEAM-X is the explanation layer built with multimodal large language models. Given a query image, a positive satellite image, a negative satellite image, and a textual instruction, GLEAM-X asks an MLLM to predict which pair is the correct match and to generate a human-readable explanation for its decision. GLEAM-X uses GLEAM-C’s matched and mismatched pairs to construct datasets and define the CVGL portion of the task, but is architecturally separate; the data identify it as being based on Qwen2.5-VL-3B-Instruct. In this division of labor, GLEAM-C performs matching and GLEAM-X explains why.

Several limitations are stated explicitly. The first is satellite-centric alignment: all modalities are aligned only through satellite imagery, which assumes high-quality, dense satellite coverage and does not explicitly model direct relationships such as UAV-to-ground photo. The second is dataset dependency and domain gaps, since training relies on VIGOR, University-1652, SetVL-480K, and MAP, and generalization to new cities, rural areas, weather conditions, or unseen sensor types is not explicitly studied. The third is the inability to use GPS-informed hard negative sampling because GPS metadata are missing in some datasets. The fourth is the absence of explicit geometric or semantic alignment mechanisms such as polar transforms, semantic segmentation, or region-level transformers. Finally, adding modalities such as thermal or LiDAR projections would require new paired datasets and possibly more careful curriculum design [2509.07450].

The future directions listed or implied in the source include incorporating more modalities and datasets, exploring GPS-augmented sampling when metadata are available, combining GLEAM-C embeddings with downstream navigation and planning tasks, and integrating additional structural or geometric inductive biases while retaining the unified backbone. Taken together, these indicate that GLEAM-C is positioned as a foundational CVGL model in the specific sense articulated by the paper: not as a universal geo-localization model in general, but as a satellite-anchored, contrastive backbone that unifies multiple views and modalities within one retrieval framework.

Source: https://www.emergentmind.com/topics/gleam-c