UBIPr Database: Biometric Data & Publication Index

• UBIPr Database is a dual-purpose resource, representing both a high-resolution periocular image dataset captured at multiple distances and a bibliographic publication index for interferometry.
• The periocular dataset features controlled capture conditions using a Canon EOS 5D DSLR, 86 subjects, and five distinct stand-off distances, enabling rigorous intra- and inter-distance verification.
• Advanced CNN approaches using fusion techniques achieve state-of-the-art results (1.31% EER), while bibliographic ambiguity poses challenges for citation and retrieval.

Searching arXiv for papers mentioning the UBIPr database and closely related terminology. The term UBIPr Database is used in the supplied literature in two distinct ways. In biometric research, it denotes a visible-spectrum periocular image dataset acquired at multiple stand-off distances and used for periocular verification experiments, including cross-distance evaluation (Alonso-Fernandez et al., 30 Oct 2025). Separately, one source explicitly states that a search for the “UBIPr Database” corresponds to the OLBIN publication database, an optical long baseline interferometry bibliography hosted by the Jean-Marie Mariotti Center (JMMC) and integrated with the OLBIN web site (Malbet et al., 2010). The dominant technical usage in the provided data is the periocular dataset, for which the available description is substantially richer; however, the naming ambiguity is itself part of the record and affects citation, retrieval, and interpretation.

1. Nomenclature and scope

In the biometric usage documented here, the UBIPr database is a periocular dataset comprising visible-spectrum periocular images captured with a CANON EOS 5D DSLR under cooperative acquisition conditions (Alonso-Fernandez et al., 30 Oct 2025). The dataset is characterized by multiple acquisition distances, specifically 4 m, 5 m, 6 m, 7 m, and 8 m, which induce controlled scale and resolution variation. The study summarized in the source uses 86 subjects, two sessions per subject, and, for each distance, one image per eye and per session, yielding 344 images per distance and 1,718 frontal periocular images in total (Alonso-Fernandez et al., 30 Oct 2025).

A separate usage appears in the interferometry literature. There, the source states that if one is searching for the “UBIPr Database,” it corresponds to the OLBIN publication database, described as an up-to-date, curated bibliography for the optical long baseline interferometry community, hosted and maintained by JMMC and linked to the OLBIN web site (Malbet et al., 2010). This resource is a publication database rather than an image dataset. The coexistence of these two usages suggests a bibliographic ambiguity rather than a shared technical object.

2. Dataset composition and acquisition conditions

The periocular UBIPr database, as used in the 2025 study, is defined by a controlled yet nontrivial acquisition setup. The imagery is high-resolution, with the paper noting 22.3 MP for the CANON EOS 5D DSLR, and the subjects are cooperative (Alonso-Fernandez et al., 30 Oct 2025). The images are frontal periocular images, and the iris is explicitly kept unmasked, reflecting conditions in which iris segmentation may be unreliable (Alonso-Fernandez et al., 30 Oct 2025).

The dataset includes five stand-off distances:

Attribute	Value	Source
Subjects	86	(Alonso-Fernandez et al., 30 Oct 2025)
Sessions per subject	2	(Alonso-Fernandez et al., 30 Oct 2025)
Distances	4 m, 5 m, 6 m, 7 m, 8 m	(Alonso-Fernandez et al., 30 Oct 2025)
Images per distance	344	(Alonso-Fernandez et al., 30 Oct 2025)
Total images used	1,718	(Alonso-Fernandez et al., 30 Oct 2025)

For each distance, the image count is computed in the source as 86 × 2 eyes × 2 sessions = 344 images (Alonso-Fernandez et al., 30 Oct 2025). This structure supports both intra-distance verification and inter-distance verification, allowing controlled analysis of performance degradation as acquisition distance changes.

The environmental variability reported in the source is limited but still relevant. Glasses are present in some images, and the paper notes that heatmap analysis indicates models tend to ignore eyeglass frames and focus on more consistent periocular skin regions (Alonso-Fernandez et al., 30 Oct 2025). Because the capture is cooperative and high quality, the source indicates reduced motion blur and reduced severe occlusions relative to more operational scenarios.

3. Region of interest definition and normalization

The periocular UBIPr protocol described in the source uses a geometry-driven normalization procedure. The authors perform manual annotation of inner/outer sclera boundaries to obtain a robust sclera radius $R_s$ (Alonso-Fernandez et al., 30 Oct 2025). This is then used for distance-specific normalization, in which images in each distance group are resized to match the average $R_s$ of that group in order to equalize scale (Alonso-Fernandez et al., 30 Oct 2025).

Cropping is performed using a $7.6 R_s \times 7.6 R_s$ square centered on the sclera center (Alonso-Fernandez et al., 30 Oct 2025). For orientation consistency, left-eye crops are horizontally flipped so that the nose is on the left, and both eyes are treated as the same identity (Alonso-Fernandez et al., 30 Oct 2025). The source emphasizes that the full periocular region is retained and that the iris is not masked.

These preprocessing choices are central to how the dataset is operationalized in later verification experiments. This suggests that reported performance depends not only on the raw image collection but also on a standardized normalization and cropping pipeline designed to reduce scale and orientation variability while preserving realistic periocular context.

4. Verification protocol and evaluation structure

The source provides a detailed verification protocol on UBIPr (Alonso-Fernandez et al., 30 Oct 2025). Two scenarios are defined: intra-distance and inter-distance verification.

For intra-distance evaluation, each distance $D_i$ is matched against itself. Genuine comparisons are computed as session 1 vs session 2, same subject, both eyes cross-compared, producing 4 comparisons per user and therefore 344 genuine scores per distance. Impostor comparisons are formed as session 1 eyes vs session 2 eyes of all other users, yielding 86 × 85 × 4 = 29,240 impostor scores per distance (Alonso-Fernandez et al., 30 Oct 2025).

For inter-distance evaluation, the source defines 10 ordered pairs with $D_i$ vs $D_j$, $i \neq j$. Genuine scores are computed as session 1 at $D_i$ vs both sessions at $D_j$, with 8 comparisons per user, giving 688 genuine scores per inter-distance pair. The impostor count remains 29,240 impostor scores per distance pair (Alonso-Fernandez et al., 30 Oct 2025).

Across all 15 distance combinations, the source reports totals of 8,600 genuine and 438,600 impostor scores (Alonso-Fernandez et al., 30 Oct 2025). Performance is summarized using Equal Error Rate (EER).

The similarity functions reported are cosine similarity and chi-square distance. The source gives the definitions

$s_{\cos}(x, y) = \frac{x \cdot y}{\|x\|_2 \|y\|_2}$

and

$R_s$0

The paper reports both measures at dataset level, while per-distance analysis focuses on ImageNet initialization with $R_s$1, stated to be consistently best (Alonso-Fernandez et al., 30 Oct 2025).

5. Use in CNN-based periocular verification

The most detailed arXiv treatment in the supplied material uses UBIPr to study complementarity and explainability in CNNs for periocular verification across acquisition distances (Alonso-Fernandez et al., 30 Oct 2025). Three architectures of increasing complexity are trained on a large periocular corpus derived from VGGFace2: SqueezeNet, MobileNetv2, and ResNet50.

The source reports the following architectural descriptions and training setup (Alonso-Fernandez et al., 30 Oct 2025):

• SqueezeNet: 18 layers, approximately 1.24M parameters, with batch normalization added between conv and ReLU.
• MobileNetv2: 53 layers, approximately 3.5M parameters, using inverted residuals and depth-wise separable convolutions.
• ResNet50: 50 layers, approximately 25.6M parameters, with residual bottlenecks.

All three are implemented in MATLAB R2024b and adapted to accept 113×113 inputs by changing the first convolution stride from 2 to 1 (Alonso-Fernandez et al., 30 Oct 2025). Input normalization is subtract 127.5 and divide by 128. Identity templates are taken from the Global Average Pooling layer immediately before the classifier (Alonso-Fernandez et al., 30 Oct 2025).

Two initialization regimes are compared: ImageNet initialization and “Face” initialization, where the model is first trained on MS-Celeb-1M (RetinaFace-cleaned MS1M; 5.1M images, 93.4K classes) and then fine-tuned on VGGFace2 (Alonso-Fernandez et al., 30 Oct 2025). Large-scale periocular pretraining uses more than 1.9M ocular crops derived from VGGFace2, specifically 1,907,572 ocular crops obtained from 953,786 valid faces, or approximately 221 crops/identity on average (Alonso-Fernandez et al., 30 Oct 2025).

The training objective is ocular identification via cross-entropy loss on VGGFace2 crops, using SGDM, batch size 128, and a staged learning-rate schedule 0.01 → 0.005 → 0.001 → 0.0001, reduced when validation plateaus, with 2% validation per user (Alonso-Fernandez et al., 30 Oct 2025).

6. Reported performance, fusion, and explainability

On pooled intra- and inter-distance trials, the source reports that ResNet50 is the best single model and that $R_s$2 consistently outperforms cosine for each backbone (Alonso-Fernandez et al., 30 Oct 2025). The dataset-level EER values are:

Model / setup	Cosine EER	$R_s$3 EER
SqueezeNet, ImageNet	5.44%	4.93%
MobileNetv2, ImageNet	2.12%	2.10%
ResNet50, ImageNet	1.73%	1.66%
SqueezeNet, Face	5.97%	5.45%
MobileNetv2, Face	2.24%	2.15%
ResNet50, Face	1.95%	1.93%

The source also states that ImageNet initialization outperforms “face” initialization across all networks and metrics, attributing this to more generic features that adapt better to periocular data (Alonso-Fernandez et al., 30 Oct 2025).

Score-level fusion is performed with linear logistic regression,

$R_s$4

with $R_s$5 (Alonso-Fernandez et al., 30 Oct 2025). Under ImageNet + $R_s$6, the best pooled result is obtained by fusing all three networks, achieving 1.31% EER, which the source describes as a new state-of-the-art on UBIPr under the paper’s protocol (Alonso-Fernandez et al., 30 Oct 2025). Relative to the best single network at 1.66%, this corresponds to a −21.14% relative improvement in EER (Alonso-Fernandez et al., 30 Oct 2025).

Per-distance behavior is also specified. For intra-distance verification, performance is stable at 4–7 m, while the 8 m condition is the most challenging (Alonso-Fernandez et al., 30 Oct 2025). For inter-distance verification, accuracy degrades with larger distance gaps, most noticeably for SqueezeNet, whereas ResNet50 remains robust with EERs under 2% even for 4 m gap (Alonso-Fernandez et al., 30 Oct 2025). The all-network fusion typically achieves EER < 1.5% and sometimes < 1% across most intra- and inter-distance scenarios (Alonso-Fernandez et al., 30 Oct 2025).

The same study uses LIME heatmaps and Jensen–Shannon divergence (JSD) to analyze complementarity (Alonso-Fernandez et al., 30 Oct 2025). The reported qualitative pattern is that MobileNetv2 attends more locally, often under the lower eyelid, whereas ResNet50 and SqueezeNet show broader patterns involving upper eyelid, sclera, and tear duct. The pupil/iris receives less attention, indicating reliance on periocular context rather than iris texture (Alonso-Fernandez et al., 30 Oct 2025). Pairwise JSD correlations are reported as 0.193, 0.292, and 0.559, supporting the conclusion that the networks focus on distinct regions and therefore benefit from fusion (Alonso-Fernandez et al., 30 Oct 2025).

7. Ambiguity, limitations, and related usage

A notable issue is terminological ambiguity. The supplied interferometry source explicitly states that the resource corresponding to the “UBIPr Database” is the OLBIN publication database, not a biometric image corpus (Malbet et al., 2010). That OLBIN resource is built on MySQL, integrated with NASA ADS, classified by approximately 70 tags grouped into seven categories, and used to generate community statistics in optical long baseline interferometry (Malbet et al., 2010). It is therefore a fundamentally different type of database. The overlap in naming is a bibliographic complication rather than a conceptual relation.

Within the biometric usage, the source notes several limitations. The principal one is potential domain shift, since models trained on VGGFace2 eye crops are evaluated on UBIPr, and generalization to lower-quality sensors and non-cooperative scenarios remains to be tested (Alonso-Fernandez et al., 30 Oct 2025). The source also reports visible sensitivity at the farthest range (8 m) and for larger inter-distance gaps, especially in weaker architectures (Alonso-Fernandez et al., 30 Oct 2025). Glasses are present in some images; although models tend to ignore the frames, attention variability around eyelashes and tear duct can still lead to divergence (Alonso-Fernandez et al., 30 Oct 2025).

The same paper outlines several future directions: evaluating on surveillance-grade sensors and non-cooperative subjects, exploring near-infrared periocular data, considering spectrum translation to build sufficient NIR training data, and integrating margin-based losses such as ArcFace together with sequential fine-tuning (MS1M → VGGFace2) (Alonso-Fernandez et al., 30 Oct 2025).

Taken together, the supplied record supports two encyclopedia-level meanings for “UBIPr Database.” In current biometric usage, it denotes a periocular verification benchmark with controlled multi-distance acquisition, strong utility for cross-distance evaluation, and a demanding protocol that has recently been used to demonstrate 1.31% pooled EER through CNN fusion (Alonso-Fernandez et al., 30 Oct 2025). In a separate bibliographic usage, the same label is mapped to the OLBIN/JMMC interferometry publication database, a curated scholarly index rather than an image dataset (Malbet et al., 2010).