CUVIRIS Dataset: ISO-Compliant Iris Recognition

• CUVIRIS dataset is a curated collection of 752 high-quality VIS iris images acquired under strict ISO/IEC 29794-6 protocols, serving as a standardized benchmark for mobile iris recognition.
• It utilizes a custom Android app with real-time quality control, including YOLOv3-Tiny for eye detection and Laplacian variance for focus evaluation, ensuring consistent image quality.
• Benchmarking with OSIRIS and IrisFormer demonstrates robust performance with high TAR and low EER, highlighting the dataset’s value for both classical and transformer-based biometric systems.

The CUVIRIS dataset is a curated collection of 752 high-quality visible-spectrum (VIS) iris images acquired under rigorous ISO/IEC 29794-6 compliance protocols using commodity smartphones. Developed to address challenges in VIS iris recognition—specifically illumination variability, pigmentation heterogeneity, and the lack of standardized mobile-capture controls—CUVIRIS provides a reproducible resource for evaluating and benchmarking both classical and modern iris recognition systems on mobile platforms. It is accompanied by a specialized Android acquisition app, state-of-the-art VIS-adapted segmentation and matching models, and open-access implementation resources, collectively supporting research toward practical, ISO-compliant smartphone-based biometrics.

1. Dataset Structure and Subject Demographics

CUVIRIS comprises 752 images collected from 47 human subjects, each contributing sixteen samples—eight per eye. Captures were performed indoors under uniform lighting, ensuring controlled presentation conditions and minimizing environmental variation.

Attribute	Value	Details
Number of images	752	High-quality, ISO-compliant
Number of subjects	47	16 images per subject; 8 per eye
Age range	18–32 years	Young adult demographics
Ethnic distribution	Diverse	25 Caucasian, 8 Hispanic, 8 Asian, 5 Black, 1 Native American
Iris pigmentation	Balanced mix	Both light and dark irides represented

Images were acquired using a Samsung Galaxy S21 Ultra mounted on a tripod, facilitating stability and session consistency. This stringent protocol supports robust recognition research by eliminating variability arising from device placement or subject movement.

2. Acquisition Methodology and Quality Control

The dataset was collected via a custom Android application developed using Android Studio. Designed for real-time field use on commodity smartphones, the application incorporates multiple mechanisms for enforcing image quality and protocol adherence:

• Frame selection and eye localization: A YOLOv3-Tiny detector, fine-tuned on UBIRIS.v1/v2, provides real-time region-of-interest framing and automatic eye detection.
• Focus and exposure control: Continuous autofocus is maintained, and images are acquired at the full sensor’s native resolution with downsampled preview mapping.
• Sharpness evaluation: The application computes the variance of the Laplacian, applying a threshold $S = 70$ for blur rejection, directly corresponding to ISO standards.
• Quality metrics verification: Ten ISO/IEC 29794-6:2015 metrics are assessed in real time using the BIQT-Iris plugin; only those frames meeting all criteria are stored.
• Metadata encoding: Filenames include structured metadata (subjectID, eye side, sessionID, trial), promoting traceability across sessions.

The acquisition process results in a frame rejection rate of ~12% and a failure-to-enroll rate under 2%, emphasizing stringent compliance with biometric quality standards and enhancing the dataset’s interoperability with ISO-compliant recognition pipelines.

3. ISO/IEC 29794-6:2015 Compliance and Dataset Integrity

CUVIRIS adheres strictly to ISO/IEC 29794-6 standards for iris image quality, which stipulate requirements across sharpness, contrast, usable area, iris/pupil ratio, and geometric constraints. Quality compliance is ensured at the point of capture, preventing the inclusion of inadequate samples and enabling meaningful cross-system performance evaluation.

This systematic approach guarantees high dataset utility for biometric evaluation and supports reproducible benchmarking in operational VIS recognition scenarios. The standardization further facilitates comparative studies with other datasets and across alternative mobile iris capture frameworks.

4. Benchmarking Performance on CUVIRIS

Performance was quantified using two distinct recognition systems:

• OSIRIS: Implements classical Daugman-style recognition via log–Gabor filter-based encoding. On CUVIRIS, OSIRIS achieved a True Accept Rate (TAR) of 97.9% at a False Accept Rate (FAR) of 0.01, with an Equal Error Rate (EER) of 0.76%.
• IrisFormer: A transformer-based matcher adapted to the VIS domain (trained on UBIRIS.v2 only), which partitioned the normalized iris into $16 \times 16$ pixel non-overlapping patches and employed a 12-layer transformer encoder. IrisFormer delivered an EER of 0.057%, demonstrating superior discrimination capacity when compared to classical methods.

A modest performance gap was observed between light-eyed and dark-eyed subsets (with dark irides yielding slightly higher EER for OSIRIS), though controlled VIS acquisition reduced the typical impact of pigmentation variation. These findings affirm that rigorous ISO-compliant capture and tailored VIS processing can effectively mitigate long-standing modality-specific challenges.

5. On-device Segmentation and Matching: LightIrisNet and IrisFormer

Technological innovation in the CUVIRIS pipeline centers on real-time, mobile-efficient segmentation and matching:

• LightIrisNet: A lightweight, multi-task segmentation network based on MobileNetV3, engineered for on-device computation. Outputs include iris and pupil masks, edge maps, signed distance transforms (SDTs), and ellipse parameters. Ellipse regression is performed via the normalized parametric vector:

$$\theta = \left(\frac{c_x}{W}, \frac{c_y}{H}, \frac{r_x}{W}, \frac{r_y}{H}, \sin\alpha, \cos\alpha\right)$$

where $(c_x, c_y)$ is the center, $(r_x, r_y)$ are radii, $\alpha$ is orientation, and $W, H$ are image dimensions.

• IrisFormer: Adapts transformer architectures to VIS iris matching. Normalized iris regions are divided into fixed-size patches and encoded via contextual patch-level embeddings and relative positional encodings (RoPE), enhancing resilience to rotational, occlusion, and glare artifacts endemic to VIS captures.

These approaches support robust, real-time segmentation and matching entirely on-device, addressing practical deployment requirements in mobile biometrics.

6. Public Release, Reproducibility, and Research Impact

A public subset of CUVIRIS (covering 10 subjects) is available via DOI, facilitating open research and benchmarking. The Android acquisition app, pretrained LightIrisNet, and VIS-adapted IrisFormer models—including code and implementation scripts—are distributed via GitHub, supporting full reproducibility and comparative experimentation.

This open-access strategy enables the research community to replicate high-quality VIS iris image acquisition, apply state-of-the-art segmentation/matching pipelines, and extend the investigation of standardized smartphone-based iris recognition.

7. Significance for Smartphone-based Biometric Applications

CUVIRIS establishes a framework for practical, ISO-compliant VIS iris recognition on widely available mobile hardware. The combination of rigorous capture protocols, validated benchmark performance across canonical and transformer-based systems, and open deployment resources offers a reference point for future progress in accessible, standards-driven biometric modalities.

The dataset’s robust composition and programmatic quality verification contribute to advancing reproducible evaluation methodologies, supporting both academic and applied research into mobile iris recognition solutions. A plausible implication is accelerated development and validation for consumer-facing biometric authentication in the VIS spectrum, with stronger interoperability guarantees enabled by strict ISO adherence.