GLOD: Multi-Domain Global-Local Approaches

• GLOD is a multi-faceted term defining distinct methodologies, including graph-level outlier/OOD detection, remote sensing object detection, diffusion-based image generation, soft-output decoding, and global-local dataset curation.
• In graph learning, evaluations reveal performance flips when repurposing binary classification datasets, emphasizing unbiased metric reporting with ROC-AUC and PR-AUC under varying downsampling scenarios.
• Advanced implementations in remote sensing and diffusion leverage transformer backbones, fusion modules, and local guidance to enhance detection precision and maintain balanced global-local information.

Searching arXiv for recent and canonical uses of “GLOD” across domains. Searching for graph-level OOD/outlier detection usages of GLOD. Searching for computer vision usages of GLOD/GLoD. In current arXiv usage, GLOD is not a single standardized term. It most prominently denotes Graph-Level Outlier Detection and Graph-Level Out-of-Distribution detection in graph machine learning, but it is also used for the Global-Local Object Detector in satellite imagery, for GRAND with log-likelihood outputs in joint detection and decoding, and, in the capitalization variant GLoD, for Global-Local Diffusion in image generation and the Global-Local Dataset for text-rich image translation (Zhao et al., 2020, Wang et al., 2024, Drapier et al., 15 Jul 2025, Sarieddeen et al., 2022, Yamada, 2024, Lu et al., 25 Feb 2026).

1. Terminological scope

The literature uses the label “GLOD” for several technically unrelated constructs. In graph mining, it refers either to graph-level outlier detection or graph-level OOD detection. In remote sensing, it names a transformer-first detector for high-resolution satellite imagery. In coding theory, it denotes a soft-output GRAND variant. In generative modeling and multimodal translation, the capitalization variant “GLoD” denotes a diffusion framework and a dataset, respectively (Zhao et al., 2020, Wang et al., 2024, Drapier et al., 15 Jul 2025, Sarieddeen et al., 2022, Yamada, 2024, Lu et al., 25 Feb 2026).

Usage	Domain	Representative source
Graph-Level Outlier Detection / Graph-Level OOD detection	Graph machine learning	(Zhao et al., 2020, Wang et al., 2024)
Global-Local Object Detector	Remote sensing object detection	(Drapier et al., 15 Jul 2025)
GRAND with log-likelihood outputs	Channel decoding	(Sarieddeen et al., 2022)
Global-Local Diffusion	Text-to-image generation	(Yamada, 2024)
Global-Local Dataset	Text Image Machine Translation	(Lu et al., 25 Feb 2026)

A common source of confusion is the proximity between GLOD and GOLD. They are distinct: GOLD is the NASA Global-scale Observations of the Limb and Disk ultraviolet imaging spectrograph, whereas GLOD/GLoD in the works discussed here denotes other concepts entirely (Solomon1 et al., 2019).

2. GLOD in graph machine learning

In the 2020 graph-mining formulation, Graph-Level Outlier Detection (GLOD) is defined over a database of graphs $\mathcal{G}=\{G_1,\dots,G_N\}$, where each graph is $G=(V,E,X)$ with nodes $V$, edges $E$, and node features $X\in\mathbb{R}^{n\times d}$. The task is to find graphs whose distributional mechanism deviates from that generating the majority of graphs. This differs from node- or edge-level anomaly detection because the entire graph is the scored unit (Zhao et al., 2020).

A common evaluation protocol repurposes binary graph classification datasets. If the original classes are $C_0$ and $C_1$, one class is selected as inlier, the other is downsampled by a fraction $\alpha\in(0,1]$, and the combined set is evaluated as an outlier-detection benchmark. The study emphasizes that a fair evaluation repeats this in both directions—$C_0$ as inlier and $C_1$ downsampled, then vice versa—and averages over multiple random downsamplings. It also distinguishes ROC-AUC, defined as $G=(V,E,X)$0, from PR-AUC, which is more sensitive under heavy imbalance (Zhao et al., 2020).

The later UB-GOLD benchmark broadens the term toward Graph-Level Out-of-Distribution detection, with unlabeled ID training graphs $G=(V,E,X)$1 drawn from $G=(V,E,X)$2 and test graphs drawn from a mixture of $G=(V,E,X)$3 and $G=(V,E,X)$4. The detector outputs an OOD score $G=(V,E,X)$5, and the graph-level decision rule is

$G=(V,E,X)$6

UB-GOLD treats both GLAD and GLOD as instances of generalized graph-level OOD detection under a single unsupervised objective (Wang et al., 2024).

3. Evaluation pathologies, benchmarks, and methodological consequences

A central empirical result in graph-level outlier detection is the performance flip observed when binary graph classification datasets are repurposed for GLOD. Across many datasets and detectors, ROC-AUC can be high when one class is downsampled yet become very low—sometimes worse than random—when the roles of the two classes are swapped. Representative values at $G=(V,E,X)$7 include, on DD, WL+LOF: $G=(V,E,X)$8 versus $G=(V,E,X)$9, PK+LOF: $V$0 versus $V$1, and OCGIN: $V$2 versus $V$3; on NCI1, the direction can reverse, as with WL+LOF: $V$4 versus $V$5 (Zhao et al., 2020).

The same study reports an approximate AUC complementarity:

$V$6

so the two AUCs often sum to approximately $V$7 when ties are negligible. It further attributes the flip to two embedding-space effects: density disparity, measured with neighborhood diagnostics such as NN-Radius, and overlapping support, quantified by NN-Disagreement%. For propagation-based methods such as the Weisfeiler–Lehman subtree kernel and the Propagation Kernel, increasing propagation depth $V$8 amplifies the gap between the two downsampling variants because sparsification and diagonal dominance increase with $V$9 (Zhao et al., 2020).

UB-GOLD reframes these issues at benchmark scale. It covers 35 datasets across four practical scenarios: intrinsic anomaly, class-based anomaly, inter-dataset shift, and intra-dataset shift, and compares 18 representative GLAD/GLOD methods. It evaluates detectors with AUROC, AUPRC, and FPR@95, and conducts multi-dimensional analyses of effectiveness, OOD sensitivity spectrum, robustness, and efficiency. Aggregate AUROC results reported in the benchmark place SIGNET at 75.86 average AUROC with average rank 3.34, followed by OCGTL at 73.14, while end-to-end methods consistently outperform two-step baselines on average (Wang et al., 2024).

Several practical consequences follow directly from these studies. The 2020 work recommends evaluating both downsampling choices, using mixtures of “X vs. Non-X” and “X vs. Y” datasets, reporting PR-AUC in addition to ROC-AUC, and tempering propagation-induced sparsification. UB-GOLD similarly emphasizes multi-metric evaluation, unsupervised threshold calibration, robustness to contaminated training data, and the fact that near-OOD detection remains substantially harder than far-OOD detection (Zhao et al., 2020, Wang et al., 2024).

4. GLOD as the Global-Local Object Detector

In remote sensing, GLOD denotes the Global-Local Object Detector, a transformer-first architecture for object detection in high-resolution satellite imagery. It replaces conventional CNN backbones with a Swin Transformer encoder, uses a convolutional neck built from UpConvMixer (UCM) blocks and Fusion Blocks, and ends with a CenterNet-inspired anchor-free head. The design targets tiny objects, severe scale variation and density, and computational constraints in overhead imagery (Drapier et al., 15 Jul 2025).

The end-to-end pipeline is defined for a fixed input of 3072×3072. The Swin backbone produces hierarchical stages S1–S4 by windowed self-attention and shifted windows. UCM blocks then upsample and refine features using asymmetric fusion, separable convolutions, CBAM attention, Highway gating, and PixelShuffle. The asymmetric fusion module is given as

$E$0

and the UCM recurrence with $E$1 repeats includes

$E$2

Fusion Blocks combine adjacent UCM outputs by channel alignment, bilinear upsampling of the lower-resolution stream, and additive fusion followed by GELU (Drapier et al., 15 Jul 2025).

The prediction head outputs center heatmaps, sub-pixel offsets $E$3, and sizes $E$4. The total loss is

$E$5

with $E$6. Training on xView uses AdamW with $E$7, $E$8, $E$9, $X\in\mathbb{R}^{n\times d}$0, cosine warm restarts, 42k training steps, per-device batch size 3, accumulated batch size 24, and gradient checkpointing. Images are resized to 3072×3072, normalized with ImageNet-1K mean and standard deviation, and trained from scratch rather than with backbone pretraining (Drapier et al., 15 Jul 2025).

Reported xView performance is 32.95 mAP50 and 28.87 mAP75. This exceeds the best CNN baseline listed, FPN-50 RFL at 28.32 mAP50, by 4.63 points, and the best transformer baseline listed, DQ-DETR at 29.56 mAP50, by 3.39 points. Ablation also attributes a substantial gain to Fusion Blocks: with FB, 32.95/28.87 mAP50/mAP75; without FB, 30.13/24.72 (Drapier et al., 15 Jul 2025).

5. GLoD as Global-Local Diffusion

In image generation, the capitalization variant GLoD denotes Global-Local Diffusion, a training-free framework for text-to-image generation that composes global prompts describing scene-level interactions with local prompts specifying object-specific attributes. The method is designed for the case where diffusion models misattribute attributes across multiple entities or ignore them when a single prompt must encode both global interactions and local details (Yamada, 2024).

GLoD uses a pre-trained diffusion model and performs multiple forward passes at each timestep to obtain unconditional, global-conditioned, and local-conditioned predicted noises. Its global guidance is

$X\in\mathbb{R}^{n\times d}$1

while its local guidance for object $X\in\mathbb{R}^{n\times d}$2 within mask $X\in\mathbb{R}^{n\times d}$3 is

$X\in\mathbb{R}^{n\times d}$4

Sampling then follows

$X\in\mathbb{R}^{n\times d}$5

The crucial design choice is that local guidance is defined relative to the base global condition inside the object mask, rather than as an independent prompt added uniformly across the image (Yamada, 2024).

The framework is implemented with Stable Diffusion v2.1, latent space $X\in\mathbb{R}^{n\times d}$6, Euler Discrete Scheduler, and training-free layout control via backward guidance using cross-attention maps and masks. Runtime scales linearly with the number of prompts because each step requires 1 unconditional + (#global) + (#local) U-Net evaluations. Bounding-box masks are central to reducing attribute leakage, and the paper defines an infection score $X\in\mathbb{R}^{n\times d}$7 to quantify undesirable cross-effects between objects (Yamada, 2024).

On a 2,500-sample controllability-oriented test set, GLoD improves the combined global-local score $X\in\mathbb{R}^{n\times d}$8 and reduces infection relative to training-free layout control. Reported values include training-free layout control at approximately $X\in\mathbb{R}^{n\times d}$9, $C_0$0, $C_0$1, $C_0$2, $C_0$3, whereas GLoD reports $C_0$4, $C_0$5, $C_0$6, $C_0$7, $C_0$8 (Yamada, 2024).

6. GLOD as GRAND with log-likelihood outputs

In coding theory, GLOD denotes GRAND with log-likelihood outputs, introduced in a soft-input, soft-output joint detection and GRAND framework. GRAND itself performs maximum-likelihood decoding by guessing additive noise sequences in decreasing order of likelihood and returning the first noise-recovered word that lies in the codebook. GLOD extends this procedure to produce bitwise log-likelihood ratios (LLRs) using Euclidean-distance metrics computed for candidate noise-recovered words (Sarieddeen et al., 2022).

For bit $C_0$9, the target quantity is

$C_1$0

The method approximates the required log-sum-exp using the small set of candidates found through GRAND’s guesswork. In its Max-Log form, it tracks the current best distance $C_1$1 and a per-bit counter-ML distance $C_1$2, yielding

$C_1$3

The paper emphasizes that the entropy of noise is typically smaller than the entropy of information bits, so a relatively small number of guesses often suffices to obtain reliable LLRs for critical bits (Sarieddeen et al., 2022).

GLOD is also embedded in an iterative turbo-GRAND loop. Extrinsic LLRs produced in one iteration are reused as a priori information to reorder noise guesses in the next iteration, improving both detection and decoding. Simulations for BCH [127,113] in AWGN and Rayleigh fading show that a few turbo-GRAND iterations match the performance of ML-detection-based soft-GRAND while average complexity grows linearly rather than exponentially with the number of symbols. The reported gains include more than 6 dB SNR advantage at BLER $C_1$4 in Rayleigh fading over schemes lacking quality soft information, with two iterations capturing most of the improvement (Sarieddeen et al., 2022).

7. GLoD as the Global-Local Dataset

In multimodal translation, GLoD stands for Global-Local Dataset, introduced as the training substrate for GLoTran in high-resolution, text-rich Text Image Machine Translation (TIMT). A GLoD sample couples a low-resolution global image $C_1$5 with a set of local slices $C_1$6 obtained by detecting, grouping, and cropping text regions from the original image, along with source-language text, target-language translation, bounding boxes, reading order, and grouping metadata (Lu et al., 25 Feb 2026).

The dataset is reported as 517,354 global-local image–text pairs across 517,354 images, spanning more than 40 real-world scenarios and 5 languages. Scenarios explicitly include documents, posters, menus, road signs, receipts, photos, and other text-rich settings with irregular layouts, small and scattered text, handwriting, decorative fonts, and non-text distractors. The curation pipeline comprises scene conceptualization, data collection and pre-filtering, text region detection and grouping, global-local translation, and quality control. An intermediate stage initially yields approximately 2M global-local pairs before rigorous filtering (Lu et al., 25 Feb 2026).

Its formal structure is tied to the global-local translation model. Starting from an image $C_1$7, a detector produces bounding boxes $C_1$8, each local crop is $C_1$9, an ordering function yields $\alpha\in(0,1]$0, and a merge operator gives $\alpha\in(0,1]$1 with $\alpha\in(0,1]$2. GLoTran then consumes the ordered local slices together with the global image and special tokens $\alpha\in(0,1]$3 and $\alpha\in(0,1]$4, using a replay window of previous translations to maintain consistency. The training objective is standard cross-entropy over target tokens:

$\alpha\in(0,1]$5

On MCiTon, GLoTran with a Qwen3-VL 8B backbone achieves an average of 50.2 BLEU / 82.2 COMET, compared with 46.7 / 80.2 for the same backbone without GLoD-based dual-perception tuning (Lu et al., 25 Feb 2026).

Across these usages, the name GLOD/GLoD consistently signals a global-local or graph-level detection perspective, but the underlying objects, assumptions, and metrics are domain-specific. In graph learning it concerns graph abnormality under unsupervised evaluation; in satellite imagery it denotes a transformer-first detector for dense multi-scale scenes; in diffusion it is a prompt-composition mechanism; in coding it is a soft-output decoding procedure; and in TIMT it is a dataset that operationalizes global-local dual perception (Zhao et al., 2020, Wang et al., 2024, Drapier et al., 15 Jul 2025, Yamada, 2024, Sarieddeen et al., 2022, Lu et al., 25 Feb 2026).