---
title: 'UFPR-VeSV: Multi-domain Benchmark Overview'
url: https://www.emergentmind.com/topics/ufpr-vesv
type: topic
---

# UFPR-VeSV: Multi-domain Benchmark Overview

Searching arXiv for UFPR-VeSV and related naming variants to ground the article in current papers.
{"query":"UFPR-VeSV OR VESV OR UFPR-Periocular site:arxiv.org", "max_results": 10}
UFPR-VeSV is not a single stable identifier in the current arXiv literature. The designation most explicitly names a 2026 surveillance dataset for unified fine-grained vehicle classification and automatic license plate recognition introduced by Lima et al. [2604.05271]. In other papers, however, “UFPR‑VeSV” appears as a naming variant of VESV, a multimodal TikTok short-video fake-news dataset [2509.15578], or as a shorthand mapping to the visible-spectrum UFPR-Periocular dataset used in ocular-biometrics fairness analysis [2110.01641]. The associated tasks, labels, acquisition pipelines, and evaluation protocols differ substantially, so the term requires domain-specific disambiguation.

## 1. Nomenclature and referential scope

The label has three documented uses in the supplied arXiv record. In vehicle surveillance, UFPR-VeSV is the dataset name used in the title and benchmark definition of Lima et al. In short-video misinformation research, the 2025 paper introduces VESV and explicitly notes that some contexts refer to the dataset as VeSV or UFPR‑VeSV. In ocular biometrics, the fairness study uses the exact dataset name UFPR-Periocular and maps it to “UFPR-VeSV” only as a visible-spectrum shorthand.

| Usage in literature | Domain | Canonical dataset described |
|---|---|---|
| UFPR-VeSV | Vehicle surveillance | FGVC + ALPR benchmark |
| UFPR‑VeSV / VeSV | Short-video misinformation | VESV |
| UFPR-VeSV as shorthand | Ocular biometrics | UFPR-Periocular (2020) |

This ambiguity is operationally important because the three resources differ in scale, modality, supervision, and intended use. One is a surveillance-grade vehicle dataset with license plate text and geometry; one is a multimodal TikTok veracity benchmark; and one is a visible-spectrum periocular dataset captured with mobile devices.

## 2. UFPR-VeSV as a unified vehicle-surveillance benchmark

In its most explicit and dataset-defining use, UFPR-VeSV is a surveillance-grade benchmark designed to unify fine-grained vehicle classification and automatic license plate recognition within one resource [2604.05271]. It comprises 24,945 images of 16,297 unique vehicles, with each unique vehicle identified by its license plate. The dataset contains 16,297 license plates with two layouts: 5,171 Brazilian and 11,126 Mercosur. Four FGVC attributes are fully annotated and validated: 13 colors, 26 makes, 136 models, and 14 types.

The dataset was collected from the Military Police of Paraná surveillance system in a single municipality, from roadside cameras on highways, urban streets, and rural roads; about 3% of images were manually captured by officers during monitoring operations. Capture conditions are explicitly unconstrained: heterogeneous viewpoints and distances, partial occlusions, nighttime infrared imaging, varying lighting, motion blur, overexposure from headlights, and multi-vehicle scenes. The final corpus includes 13,842 rear-view and 11,103 frontal-view images, with 5,372 nighttime infrared images, corresponding to 21.5% of the dataset.

Construction involved a staged filtering pipeline. From 30,240 collected images, 1,253 were discarded for extreme license-plate occlusion, severe degradation, or poor framing, and 4,042 highly similar samples were removed, yielding the final 24,945 images. YOLOv11 was used to detect vehicles and standardize framing, with manual corrections for multi-vehicle scenes and difficult cases. Some images had pre-existing green borders; a 5-pixel margin was removed uniformly after verifying that license plates were unaffected.

The label space includes domain-specific handling for difficult cases. Infrared images form a dedicated “infrared” color class. A “multicolored” class covers vehicles with no predominant color. Due to severe occlusions or limited visible cues, motorcycles, scooters, and rear-view truck-like vehicles can have color, make, and model set to “unknown” when the main body is obstructed. Makes and models with fewer than 25 samples are merged into “others,” and underrepresented colors are merged into closest hues such as purple to blue, garnet to red, and gold to beige.

## 3. Annotation, protocol, and integrated evaluation

UFPR-VeSV’s distinctive feature is joint validation of visual attributes and license-plate information [2604.05271]. License-plate text was manually annotated and used to retrieve official vehicle records via SENATRAN in order to validate color, make, model, and type. License-plate corner coordinates were obtained in two stages: YOLOv11 detects the license-plate region, and CDCC-NET regresses corner points. A matching algorithm associates plates to the main vehicle bounding box, followed by manual review of misdetections and text-like false positives. If official data conflicted with image appearance, the license-plate text was revisited, re-annotated, and the attributes were re-fetched and verified.

The dataset also provides viewpoint labels, camera capture mode as day or IR, and privacy protection through RetinaFace-based blurring of drivers’ and pedestrians’ faces with manual quality control. In Brazil, license plates are not personal data, but faces were anonymized to meet ethical norms.

The recommended evaluation uses five non-overlapping folds with two constraints: all images from the same license plate remain in a single fold, and multi-attribute stratification preserves per-class distributions across color, make, model, and type. Ten deterministic train/validation/test splits at ratio 3:1:1 are generated from the five folds. Each fold serves as test exactly twice; for test fold index $t$, the validation folds are $(t+1) \bmod 5$ and $(t+2) \bmod 5$, and the remaining three folds form the training set.

Benchmarking covers both FGVC and ALPR. In end-to-end fine-tuning, EfficientNet-V2 is the strongest single-task FGVC baseline across tasks, reaching 93.5% micro-accuracy, 71.5% macro-accuracy, and 73.8% F1 for color; 96.1%, 89.0%, and 90.2% for type; 94.4%, 85.0%, and 86.4% for make; and 90.9%, 86.2%, and 87.3% for model. Top-$k$ results are also high with EfficientNet-V2, including color top-1/top-2/top-3 micro-accuracy of 93.5% / 98.0% / 99.1% and macro-accuracy of 71.5% / 87.0% / 93.5%. For ALPR, ParSeq-Tiny attains 98.0% license-plate-level accuracy and 99.7% character-level accuracy, outperforming GP-ALPR.

The paper also studies joint use of FGVC and ALPR. Using the best methods, simultaneous correctness drops as attributes accumulate: 96.1% for type alone, 91.5% for type plus make, 85.5% for type plus make plus model, and 80.2% for type plus make plus model plus color. For integration with ALPR, the strictest conjunction of LPR plus make, model, type, and color yields Validation Rate 81.1%, Conflict Rate 18.9%, and Recovery Rate 68.4%. Error analysis identifies infrared capture as the dominant failure source: although IR accounts for 21.5% of the dataset, it contributes 53.4% of make, 42.9% of model, and 39.4% of type top-1 errors.

## 4. Long-tailed vehicle color recognition on UFPR-VeSV

A subsequent study treats UFPR-VeSV as a challenging long-tailed benchmark for vehicle color recognition under severe class imbalance [2606.13625]. In that setting, the dataset retains 24,945 images of 16,297 unique vehicles and uses 13 color classes: white, black, silver, gray, red, blue, green, yellow, orange, brown, beige, multicolored, and unknown. The class distribution is explicitly long-tailed. The most frequent class, white, has 7,381 images, while the rarest, brown, has only 34. Head classes are white, unknown, silver, black, and red; the tail includes blue, multicolored, green, beige, yellow, orange, and brown.

The study emphasizes macro-level evaluation because micro accuracy can remain high even when minority classes are poorly recognized. It reports
$$
A_{\text{micro}}=\frac{1}{N}\sum_{j=1}^{N}\mathbf{1}[\hat{y}_j=y_j]
$$
and
$$
A_{\text{macro}}=\frac{1}{C}\sum_{c=1}^{C}\frac{1}{|S_c|}\sum_{j\in S_c}\mathbf{1}[\hat{y}_j=y_j].
$$
Training uses UFPR-VeSV’s official stratified 3:1:1 train–validation–test split, repeated as ten independent splits, with evaluation on real test images only.

Two synthetic minority-class augmentation strategies are studied. The first uses text-conditioned image generation with RunDiffusion/Juggernaut-XL; 19,144 images were generated and 12,951 retained after manual quality control and duplicate removal. The second uses image-conditioned color editing with Gemini 2.0 Flash; 14,040 images were generated and 6,028 retained after quality control. Two independent annotators judged realism, correct target color, and domain plausibility, with Cohen’s $\kappa=0.43$ for RunDiffusion and $\kappa=0.41$ for Gemini. Moderate agreement is attributed to ambiguous color boundaries such as gray versus silver and beige versus brown, as well as low-light surveillance imagery.

The best-performing system combines Gemini-augmented training, weighted cross-entropy, Linear Warmup with Cosine Decay, color-safe augmentation, foreground-aware preprocessing with SAM 2, and hard-voting ensemble fusion. It reaches 94.6% micro accuracy, 79.7% macro accuracy, and 78.8% macro F1, improving macro accuracy by 8.2 percentage points over the prior benchmark and by 7.2 percentage points over the strongest single-model baseline in that study. The strongest single-model baseline without synthetic data is DINOv3-Large at 94.0% micro and 72.5% macro. Gemini augmentation alone contributes +2.4 percentage points macro, and the paper finds that image-conditioned edits are more beneficial than larger volumes of text-to-image synthesis because they preserve surveillance context.

Manual error analysis on three representative folds examined 785 errors. Of these, 459, or 58.5%, were labeled as inherently ambiguous, while 326, or 41.5%, were labeled as correctable model errors. Based on that split, the study estimates an upper bound of about 96.8% micro and 88.5% macro accuracy under the given capture conditions. This places a substantial fraction of residual error in the image formation process itself rather than in classifier capacity alone.

## 5. UFPR‑VeSV / VeSV as a short-video fake-news dataset

In a different research line, UFPR‑VeSV appears only as a naming variant of VESV, a short-video fake-news dataset introduced for TikTok veracity analysis [2509.15578]. VESV contains 603 TikTok videos, with 324 real and 279 fake, and durations from 10 seconds to 10 minutes. The paper states that videos are categorized as Fake, Real, or Ambiguous, following prior benchmarks, but the exact count of ambiguous items is not separately reported; in per-class reporting, Ambiguous is grouped with Fake.

VESV was created to address shortcomings of prior short-video datasets such as FakeTT. It covers TikTok videos published in 2020 or later and broadens topical scope to COVID‑19, climate change, cancer, technology, and others. The dataset is multimodal by construction. Video is processed as RGB frames sampled at 3 fps and cropped to 224×224. Audio is provided as raw audio and Mel-spectrograms, with modeling based on CLAP audio features over 8-second segments. Textual information includes standardized hashtags and captions, usernames and URLs expanded into sentences, speech transcribed using Wav2Vec2 with manual edits, and on-screen text plus relevant screenshot content documented as thoroughly as possible.

Annotation was performed by a team of three annotators following a multi-step process, with assistance and verification by linguistic experts. Two annotators independently labeled each video, and disagreements were reviewed and resolved by the first author. The dataset is evaluated with a repeated 6-fold protocol: in each run, data are split into six parts with one part for testing, one for validation, and the rest for training; this is repeated three times with different random splits and averaged.

The associated model, Heterogeneous Fusion Net, combines a Decision Network for dynamic modality weighting with a Weighted Multi-Modal Feature Fusion module. On VESV, ablations show that generic fusion is weaker than WMFF, and that DecisionNet improves WMFF further: concatenate reaches 73.42 Macro F1, add 71.66, cross-attention 74.40, WMFF without DecisionNet 77.55, and full HFN 80.33 with 80.39 accuracy. Relative to the previous best method, FakingRecipe, the reported improvement on VESV is +4.14% Macro F1 and +4.02% Accuracy. The paper does not provide a download URL, code repository link, or licensing terms for VESV.

## 6. UFPR-VeSV as a shorthand for UFPR-Periocular

A third usage appears in visible-spectrum ocular biometrics, where a fairness study maps “UFPR-VeSV” to the UFPR-Periocular dataset released in 2020 and captured with mobile RGB devices in unconstrained scenarios [2110.01641]. In that context, the exact dataset name is UFPR-Periocular rather than UFPR-VeSV. The corpus contains 33,660 ocular samples from 1,122 subjects, acquired by 196 different mobile devices, with both left and right ocular regions available. Subjects are 18 to 79 years old, and the study defines young adults as 18–39, middle-aged adults as 40–59, and older adults as 60–79.

The fairness evaluation uses subject-disjoint, gender- and age-balanced splits rather than an official fixed partition. For user verification, training uses 780 participants and testing uses 342. The study fine-tunes ResNet-50, MobileNet-V2, ShuffleNet-V2, and EfficientNet-B0. Verification reports Equal Error Rate and FNMR at fixed FMR values of 0.01 and 0.1. Across models, mean EER is 8.04% for young adults, 9.01% for middle-aged adults, and 9.88% for older adults. At the same time, older adults often exhibit lower FNMR at the fixed FMR operating points used in the paper. For example, with ResNet-50 and left-right fusion, EER is 9.06% for young, 8.85% for middle-aged, and 11.01% for older adults, while FNMR@FMR=0.01 is 36.49%, 27.90%, and 17.41%, respectively.

Gender classification is reported as high and largely equivalent across age groups and genders, with a negligible male–female gap of about 0.09% on average. Age-group classification is more difficult: young adults are easiest, while middle-aged and older adults are harder and fairly similar. Average exact and 1-off accuracies are 46.37% and 83.14% for young adults, 27.5% and 69.45% for middle-aged adults, and 24.1% and 64.5% for older adults. The paper therefore uses the dataset to argue for largely equivalent performance of ocular biometrics across gender and age groups in verification and gender classification, while also noting age-related differences in specific operating regions and age prediction tasks.

Taken together, these three usages show that “UFPR-VeSV” is less a single dataset identity than a collision point between separate UFPR-associated resources. In present usage, the most direct and title-level referent is the 2026 vehicle-surveillance benchmark for unified FGVC and ALPR, while the short-video and periocular meanings arise from naming variation and shorthand rather than from a shared dataset lineage.

Source: https://www.emergentmind.com/topics/ufpr-vesv