ImmerIris: A Large-Scale Dataset and Benchmark for Off-Axis and Unconstrained Iris Recognition in Immersive Applications

AI summary

Overview

Research area: Computer vision and biometrics — specifically iris recognition for immersive applications such as extended reality (XR) and virtual reality (VR).

Technical level: Intermediate. The paper is readable by anyone familiar with basic biometric recognition and deep learning, though the benchmark tables assume some knowledge of verification metrics (false rejection rate at a given false acceptance rate) and identification accuracy.

One-sentence scope: The paper introduces ImmerIris, a public iris dataset of 499,791 ocular images from 546 subjects captured off-axis and unconstrained through a VR headset, benchmarks seven existing iris recognition methods against it, and proposes a normalization-free recognition paradigm called NormFree.

What This Paper Is About

Iris recognition has traditionally worked well when users gaze directly at dedicated cameras in controlled conditions, but immersive applications such as XR use side-mounted headset cameras that capture eyes at an angle and in unconstrained settings. This introduces perspective distortion, variation between images of the same eye, and quality degradation, and no large public dataset captured all three of these challenges. The paper's goal is to fill that gap with a large dataset and evaluation protocols, and to test whether existing state-of-the-art methods still work under these conditions.

Key Contributions

1. The ImmerIris dataset. A large-scale, off-axis and unconstrained iris dataset acquired via VR headsets, containing 499,791 ocular images from 546 subjects. The authors state it is, to their knowledge, the largest public iris dataset to date and among the first dedicated to immersive applications.

2. A comprehensive benchmark. Eight evaluation protocols organized to isolate and combine the effects of perspective distortion, intra-subject variation (illumination, gaze), and quality degradation (occlusion, dilated pupil, specular reflection, motion blur), covering single-eye and dual-eye operation for both verification and identification.

3. Evidence that existing methods do not transfer. Seven normalization-based state-of-the-art iris recognition systems generalize unsatisfactorily from controlled to immersive data, pointing to the normalization stage as a major bottleneck.

4. A normalization-free paradigm (NormFree). A simple end-to-end design that crops the iris region with a bounding box and learns directly from minimally adjusted ocular images, outperforming normalization-based prior art in most cases. An ablation called NormKeep (same architecture, normalization retained) demonstrates the stand-alone benefit of removing normalization.

Main Findings

• Dataset scale and uniqueness: ImmerIris contains 499,791 ocular images from 546 subjects, compared with existing datasets such as CASIA-IrisV4 (20,000 images, 1,000 subjects), ND-CrossSensor (117,503 images, 676 subjects), VISOB (95,107 images, 550 subjects), and PolyU Iris DB (142,005 images, 384 subjects, the closest prior work, which is off-axis but semi-controlled). ImmerIris is the only listed dataset marked off-axis and unconstrained.

• Quality degradation is widespread: Approximately 42% of the retained images are degraded in at least one of six quality dimensions: occlusion by eyelid, occlusion by eyelash, extensively dilated pupil, extreme off-axis gaze, specular reflection, and motion blur.

• Cross-domain failure is severe: Models trained on CASIA-IrisV4-Thousand (CASIA-T) perform well on CASIA-T itself at FRR@FAR of 1e-1 (for example, Gabor 0.36%, OM 0.24%, ComplexIrisNet 1.08%), but collapse on the Immer-Any protocol (for example, Gabor 32.12%, OM 30.85%, Maxout 38.83%, ComplexIrisNet 42.25% at FAR = 1e-1). The authors attribute this primarily to the domain gap, since inputs are nominally aligned by normalization.

• More training data helps, but does not solve the problem: DNN-based SOTAs trained on ImmerIris (whose training set is 25× larger than CASIA-T) perform much better on Immer-Control than CASIA-trained models, yet all degrade sharply as protocols add variation and distortion. The best SOTA still performs unsatisfactorily on Immer-Select and Immer-Any.

• Gaze variation is the hardest isolated factor: Across isolated-challenge protocols, gaze variation accounts for an average 36.99% degradation, making it the most challenging single factor. Occlusion is only moderate, increasing FRR@FAR(1e-5) by an average of 9.34% relative to Control. Dilation does not substantially degrade recognizability, but performance drops considerably on Immer-Light, where irises naturally dilate and constrict with illumination.

• NormFree is competitive despite its simplicity: It ranks first or second in almost all cases. On dual-eye Immer-Any it reports 1.24 / 13.29 / 40.45 FRR@FAR at 1e-1 / 1e-3 / 1e-5, and on dual-eye Immer-Occlusion, Immer-Dilation, and Immer-Light it achieves the best reported results (for example, 3.92, 1.87, and 6.12 at FAR = 1e-5 respectively).

• The advantage is not decisive under gaze variation: Under gaze variation, NormFree outperforms most SOTAs but without a clear margin, which the authors identify as an area needing dedicated work.

• NormKeep lags NormFree: The normalization-retaining ablation usually falls behind the best SOTAs and shows a considerable gap from NormFree; in isolated-challenge protocols the gain over NormKeep is 4.12–16.82%.

• Ablations on scale and normalization technique: Replacing the IR-50 backbone with a smaller IR-18 leaves NormFree ahead of NormKeep on Immer-Any. Swapping in a more adaptive normalization implementation improves NormKeep only marginally, suggesting the drawback is general rather than tied to one implementation.

• Identification shows similar trends: On left-eye rank-1 accuracy, NormFree reaches 99.52 on Control, 98.91 on Fix, 93.87 on Select, 94.39 on Any, 98.23 on Occlusion, 98.36 on Light, and 95.49 on Gaze. Right-eye and remaining identification results are deferred to the supplementary material.

Methodology in Plain English

Collection. Volunteers wore a general-purpose VR headset (Skyworth Pancake XR) fitted with custom acquisition software and side-mounted cameras that produce off-axis near-infrared images. Subjects sequentially gazed at each of nine numbered red squares arranged in a 3×3 grid; at each gaze point the screen brightness stepped through 11 levels, and the headset captured 5 ocular images per eye at 640×640 resolution per level. That yields 990 images per subject for both eyes combined, and 540,540 images across 546 subjects before cleaning. Subjects were Asian adults aged 20–40 with a nearly balanced biological sex distribution, gave informed consent, received no compensation, and had no personal details recorded; the procedure was approved by an institutional review board.

Cleaning and annotation. A pretrained ocular detection model produced iris bounding boxes and 36,697 undetectable images were discarded; another 4,052 were removed by visual inspection, leaving 499,791. The remaining images were scored along the six quality dimensions and thresholded into normal versus degraded.

Splitting. The data was split 7:3 into training (347,927 images, 380 subjects) and test (151,864 images, 166 subjects), with no subject overlap, for open-set evaluation. The left and right eye of the same subject are treated as two distinct classes.

Protocols and pairing. Four protocols isolate occlusion, dilation, illumination change, and gaze variation; four more combine challenges in increasing difficulty: Immer-Control (fixed gaze, cooperative, only off-axis distortion), Immer-Fix (fixed gaze but degradation allowed), Immer-Select (extreme gaze points removed; specifically right 3/6/9 and left 1/4/7), and Immer-Any (no restriction). For verification, up to 1.5M genuine pairs and up to 3M impostor pairs are sampled per protocol so false acceptance rates down to 1e-5 can be estimated. For identification, a shared gallery holds one high-quality sample per class (except Immer-Dilation, which is omitted) and each entry is paired with up to 100 probes per class. Dual-eye testing is omitted for Immer-Select because removing extreme gaze points from both sides would leave only gaze points 2/5/8.

The proposed method. Instead of segmenting the iris, parameterizing its contour, and unwrapping it into a rectangular strip, NormFree detects the iris region with the same pretrained detector used in cleaning, crops it with a square bounding box expanded by a factor of 1.2 to keep surrounding context, and resizes it. Feature extraction uses a ResNet backbone comparable in size to the SOTAs and an angular-margin loss (ArcFace), following practices from modern face recognition rather than iris-specific preprocessing.

Why This Matters

Research impact. The paper supplies the field with a large public dataset and eight standardized protocols, which gives researchers a shared yardstick for a problem that was previously hard to study because most iris datasets are proprietary or small. It also reframes a long-standing architectural assumption — that normalization must precede feature extraction — as technical debt that deep networks may be able to absorb.

Real-world applications (based on the scenarios the paper discusses):

• Seamless user login on XR and VR headsets, where recognition happens without the user breaking immersion or performing a deliberate authentication gesture.
• In-headset payments and other sensitive transactions that need identity confirmation while the headset is worn.
• Unconstrained identification in open scenes, corresponding to the paper's Immer-Any protocol, where no cooperation requirement can be imposed on the user.
• Contrast with traditional controlled iris recognition in institutional settings such as border control, which the paper uses as the baseline scenario to show what changes in immersive use.

Industry relevance. Headset and XR hardware makers need identification that works with side-mounted cameras, imperfect inter-pupillary distance, and partial eye closure; the paper notes that data scarcity has long hindered iris recognition research, and a public dataset plus a lightweight cropping-based method lowers the barrier to building such features. The finding that normalization is fragile under realistic conditions is directly actionable for biometrics vendors whose pipelines inherit that stage.

Future Directions

• Improving gaze robustness. The authors state that NormFree lacks a decisive advantage under gaze variation and that dedicated improvements may be required, which they say they will take up in future research.
• Methodological advances beyond normalization removal. The benchmark shows that even the best existing method performs unsatisfactorily on Immer-Select and Immer-Any, so new recognition architectures or training objectives are needed for the fully unconstrained setting.
• Revisiting the normalization stage. Since a more adaptive normalization implementation only marginally helped NormKeep, future work could investigate whether normalization should be redesigned fundamentally or dropped altogether in the pipeline.
• Broadening the evaluation. Several results (right-eye identification, full supplementary analysis, implementation details) are deferred to the supplementary material, and the dataset is described as being released publicly to support future research.

Target Audience

Biometrics and iris recognition researchers; computer vision researchers working on datasets, benchmarks, and open-set recognition; XR and VR engineers building authentication into head-mounted devices; and graduate students looking for a well-documented benchmark that compares traditional, controlled-setting methods against realistic immersive conditions.

Authors’ abstract

Recently, iris recognition is regaining prominence in immersive applications such as extended reality as a means of seamless user identification. This application scenario introduces unique challenges compared to traditional iris recognition under controlled setups, as the ocular images are primarily captured off-axis and less constrained, causing perspective distortion, intra-subject variation, and quality degradation in iris textures. Datasets capturing these challenges remain limited. This paper fills this gap by presenting a large-scale iris dataset collected via head-mounted displays, termed ImmerIris. It contains 499,791 ocular images from 546 subjects, and is, to our knowledge, the largest public iris dataset to date and among the first dedicated to immersive applications. It is accompanied by a comprehensive set of evaluation protocols that benchmark recognition systems under various challenging conditions. This paper also draws attention to a shared obstacle of current recognition methods, the reliance on a pre-processing, normalization stage, which is fallible in off-axis and unconstrained setups. To this end, this paper further proposes a normalization-free paradigm that directly learns from minimally adjusted ocular images. Despite its simplicity, it outperforms normalization-based prior arts, indicating a promising direction for robust iris recognition.