Ocular Verification for Virtual Reality
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "Ocular Verification for Virtual Reality".
Jane: Here's the gist of the paper. Virtual reality (VR) headsets present novel technical challenges for reliable authentication due to unconstrained acquisition typical of VR-based data User authentication relies on biometric cues such as iris in such headsets Score-level fusion of iris with periocular cues has been proposed as a partial remedy precisely because periocular texture remains informative when the iris channel degrades The paper examines evaluating ISO/IEC 29794-61 iris quality metrics on the VRBiom dataset and analyzing their limitations They observe that some metrics (e.g., margin adequacy) fail on VR-acquired data; whereas, image adjustments primarily benefit periocular recognition, and multimodal fusion lowers EER by ∼11% over unimodal iris recognition performance.
Tom: Evaluation of Quality Metrics The researchers rigorously evaluate twelve ISO/IEC 29794-61 iris quality metrics on the VRBiom dataset using two frameworks: MITRE-BIQTIris and University of Notre Dame (UND).
Jane: They derive statistical distributions (µ ± σ) of standard ISO/IEC 29794-6 metrics across their dataset to establish a normative baseline for future VR-based biometric studies.
Tom: The analysis reveals that the VR-based data may not satisfy currently recommended values for most of the metrics.
Title and authors: Tom: Hey everyone, let's talk about who wrote this stuff. It's Husanpreet Singh, Robert Tran, Ayushree Kharel, and Sudipta Banerjee from the University of Wyoming and San Diego State University. They are the team putting this into practice.
Jane: I noticed they mention that traditional iris recognition protocols often struggle with the unconstrained acquisition typical of VR headsets, which is a key challenge they're addressing right there in the title, "Ocular Verification for Virtual Reality".
Lu: What's interesting about their setup is that they are using a publicly available dataset called VRBiom to establish a baseline for how good those standard ISO/IEC twenty-nine thousand seven hundred ninety-four-sixty-one iris quality metrics actually are under these weird VR conditions.
Meng: From an engineering standpoint, I wonder how robust their baseline is if the data distribution in that dataset doesn't perfectly reflect real-world, deployed VR hardware and user behavior.
Lalam: They are looking at three main things: evaluating those quality metrics, fixing the data-specific challenges like off-axis gaze or bad lighting using deep learning models, and then doing both iris and periocular recognition.
The paper's summary: Tom: So, to recap what this paper is trying to do, it first looks at twelve ISO/IEC twenty-nine thousand seven hundred ninety-four-sixty-one iris quality metrics on the VRBiom dataset and finds that some of those metrics don't work well with VR data.
Jane: What does that mean practically for someone using a system? It means we can't just trust the standard quality checks anymore when dealing with data captured by a headset.
Lu: They found that metrics like Margin Adequacy, which is supposed to check how good the margins of the iris are, produce scores substantially below eighty, and they attribute that low score mainly to how the camera is set up off-axis rather than just bad image quality itself.
Meng: That's a huge data point. It suggests that sometimes what looks like poor image quality is actually just a problem with the geometry of how the VR system is capturing the eye, which means we need to fix the capture first, not just enhance the picture later.
Lalam: The paper also points out that while image adjustments help periocular recognition more than iris recognition alone, combining them—score-level fusion—can reduce error rates by about eleven percent compared to using just one channel.
The paper's improvements: Tom: Now let's talk about the actual fixes they propose. They have this three-module process that starts with Ocular Image Adjustments to tackle off-axis perspectives, specular reflections, and non-uniform illumination in VR data.
Jane: That sounds like a heavy lifting pipeline. What are these specific technical tools they are using to handle those problems?
Lu: First, they have Off-axis Correction using H8Net plus DINOv3 which uses geometric correction to estimate and mitigate the perspective distortion from the off-axis VR sensor arrays.
Meng: Okay, so correcting the geometry first is smart. But what about those shiny reflections that just wash out details?
Lalam: They address that with Specularity Removal using a frozen vision transformer backbone for an RGB-only single image highlight removal pipeline to reconstruct reflection-free images, and then Non-Uniform Illumination Restoration using UNIRNet for illumination enhancement and contrast correction.
Conclusion: Tom: So we've seen the results. They tested unimodal iris recognition versus periocular recognition separately, and then they did a multimodal weighted score-level fusion to see the final performance boost.
Jane: The best result they found was with the Fusion forty/sixty configuration, which reduced the error rate to zero point three three and boosted signal detection to zero point eight eight. That's a significant jump over unimodal iris recognition.
Lu: They concluded that this weighted score-level fusion improves overall performance by boosting the area under the curve by about sixteen percent and lowering the error rate by about eleven percent compared to just using iris recognition alone.
Meng: It’s interesting because it shows that relying on periocular texture alone substantially outperforms unimodal iris verification across unconstrained VR feeds, which makes sense given how resilient those structures are to reflections.
Lalam: The authors also suggest that future work should look at designing an end-to-end pipeline for restoration and they voiced a concern that overly complex restoration networks might actually hurt verification by over-smoothing the fine iris patterns.
Tom: So, the final word on "Ocular Verification for Virtual Reality" is that combining those channels with smart adjustments provides a solid boost to performance in VR environments, but we still have to watch out for over-processing artifacts.
University of Wyoming · San Diego State University
cs.CV
Submitted: 2026-07-22
Updated: 2026-10-30
Code: https://github.com/mitre/biqt-iris
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
Importance score: 87/100
The gist: Virtual reality (VR) headsets present novel technical challenges for reliable authentication due to unconstrained acquisition typical of VR-based data User authentication relies on biometric cues…
Key concepts
- ISO/IEC 29794-61 Metrics
- These are twelve standardized measurements used to rigorously evaluate the quality of an iris image. The researchers found that VR data often fails these standards, suggesting current quality thresholds might be too strict or not suited for VR acquisition scenarios.
- Ocular Image Adjustments Pipeline
- This three-step processing sequence corrects common VR image flaws: geometric correction fixes off-axis viewing angles, UnReflect removes glare from reflections, and UNIR-Net restores contrast from uneven illumination. This pipeline prepares the image for accurate biometric analysis.
- Score-Level Fusion
- Instead of just using one biometric cue (like iris), this technique combines the scores derived from both iris recognition and periocular recognition. By weighting these scores, the system leverages the strengths of each feature to create a more robust and accurate final verification result.
Terminology
Summary
Virtual reality (VR) headsets present novel technical challenges for reliable authentication due to unconstrained acquisition typical of VR-based data User authentication relies on biometric cues such as iris in such headsets Score-level fusion of iris with periocular cues has been proposed as a partial remedy precisely because periocular texture remains informative when the iris channel degrades The paper examines evaluating ISO/IEC 29794-61 iris quality metrics on the VRBiom dataset and analyzing their limitations They observe that some metrics (e. p. 1 Score-level fusion
of iris with periocular cues has been proposed as a partial remedy precisely because periocular texture remains informative when the iris channel degrades, but
the underlying quality assessment problem to evaluate the reliability of frame
remains largely unaddressed for VR-specific data [3]. Therefore, we need to systematically address the shortcomings of the current
quality metrics and potentially revise them keeping in mind the emerging need of
VR-based deployment. We observe that some metrics (e. , margin adequacy) fail on VR-acquired data; whereas, image adjustments
primarily benefit periocular recognition, and multimodal fusion lowers
EER by ∼11% over unimodal iris recognition performance.Where the paper says it
Where the paper says it
Evaluation of Quality Metrics
The researchers rigorously evaluate twelve ISO/IEC 29794-61 iris quality metrics on the VRBiom dataset using two frameworks: MITRE-BIQTIris and University of Notre Dame (UND). They derive statistical distributions (µ ± σ) of standard ISO/IEC 29794-6 metrics across their dataset to establish a normative baseline for future VR-based biometric studies. p. 4 Unlike traditional biometric datasets that use controlled,
frontal acquisition, VR devices present a unique set of constraints that necessitate a re-evaluation of established quality benchmarks. Our primary objective is
to empirically derive the statistical distributions (µ ± σ) of standard ISO/IEC
29794-6 metrics across our dataset (see Table 1), thereby establishing a normative baseline for future VR-based biometric studies. The results indicate that the VR-based
data may not satisfy currently recommended values for most of the metrics. Our
analysis further reveals significant divergence between algorithmic implementations based on segmentation routines and handling of invalid cases between the
two frameworks. We
omit the ISO quality evaluation results for OAC from Table 2, as the geometric
warping yielded inferior metric scores compared to the original frames. We further investigate why the Margin Adequacy (MAdeq) metric (see Table
1 for details) produces values substantially below the recommended threshold
of 80 or greater. The final MAdeq score is determined by the smallest of the four
margins, meaning the metric is constrained by the narrowest distance between
the iris and the frame boundary. Therefore, the low observed mean score is due
to off-axis camera setup rather than evidence of poor image quality.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Image Enhancement Techniques
The proposed approach involves three distinct processing modules, starting with Ocular Image Adjustments to address off-axis perspectives, specular reflection removal, and non-uniform illumination. This enhancement pipeline includes:
-
Off-axis Correction (H8Net+DINOv3) which uses geometric correction to estimate and mitigate perspective distortion. p. 3
Where the paper says it
1 Ocular Image Adjustments We address the off-axis perspectives, hardware-induced specular reflections, and uneven or non-uniform illumination in VR data using three operations. Off-axis Correction (H8Net+DINOv3): Algorithm 1 details the pipeline that uses geometric off-angle correction (co) to mathematically estimate and mitigate the perspective distortion inherent to off-axis VR headset sensor arrays.
-
Specularity Removal (UnReflect) which adopts an RGB-only single-image highlight removal pipeline using a frozen vision transformer backbone to reconstruct reflection-free image. p. 3
Where the paper says it
To remove this, we adopt UnReflectAnything [17] framework, an RGB-only single-image highlight removal pipeline. We use a frozen vision transformer backbone (DINOv3) [20] to extract multi-scale feature representations from the periocular data, while the inpainting module fills in the corrupted regions to reconstruct reflection-free image.
-
Non-Uniform Illumination Restoration (UNIR-Net) which employs a network combining illumination enhancement, attention mechanisms, visual refinement, and contrast correction. p. 4
Where the paper says it
0) 15: fcurr ← WarpPerspective(fcurr, H) ▷ Apply geometric flattening 16: end if 17: Fout ← Fout ∪ {fcurr} 18: end for 19: return Fout 20: end function illumination restoration network. The network combines illumination enhancement, attention mechanisms, visual refinement, and contrast correction to improve image quality and visibility.
Verification and Fusion Strategies
The paper performs ocular verification using unimodal schemes first:
-
Iris recognition using ArcIris. p. 3
Where the paper says it
Secondly, we perform a first-of-its kind comprehensive analysis of ISO/IEC 29794-6 Quality Evaluation in VR-based ocular images using two frameworks developed independently by MITRE and UND; see Algo 2. Thirdly, we perform Ocular Verification and Multimodal Fusion by performing iris recognition (using ArcIris), periocular recognition (using MobileFaceNet), and finally a weighted multimodal iris and periocular fusion at score level for ocular verification; see Algo 3.
-
Periocular recognition using lightweight mobilefacenetbased resnet architecture. p. 4
Where the paper says it
We utilize the procedure detailed in Algorithm 2 to systematically process our datasets, providing researchers with both the statistical spread of current VR imagery. 3 Ocular Verification and Fusion We performed ocular verification firstly using unimodal schemes: (i) iris recognition using ArcIris, a deep learning-based model submitted to NIST IREX X evaluation by UND, (ii) periocular recognition using lightweight mobilefacenetbased resnet architecture.
Subsequently, a multimodal weighted score-level fusion is performed to combine iris and periocular cues for boosting verification performance. p. 4 3 Ocular Verification and Fusion
We performed ocular verification firstly using unimodal schemes: (i) iris recognition using ArcIris, a deep learning-based model submitted to NIST IREX X
evaluation by UND, (ii) periocular recognition using lightweight mobilefacenetbased resnet architecture. Secondly, we performed a multimodal weighted score
fusion to combine iris and periocular cues for boosting recognition. Ocular Verification for Virtual Reality 7
Algorithm 2 ISO Metric Quality Evaluation and Statistical Analysis
1: Input: Evaluated Frame Dataset F ∈ {Forig, FOAC , FUR, FUNIR}, BIQT engine EBIQT ,
Set of Target ISO Metrics Q
2: Output: Dataset Statistical Summary Tstats (containing µ and σ for all q ∈ Q)
3: function EvaluateQuality(F, EBIQT , Q)
4: Tstats ← ∅
5: for each q ∈ Q do
6: Vq ← ∅ ▷ Initialize empty sequence for metric values
7: end for
8: for each frame f ∈ F do
9: R ← RunBiqt(f, EBIQT ) ▷ Generate JSON quality report
10: for each q ∈ Q do
11: v ← ParseMetric(R, q) ▷ Extract scalar quality value
12: Vq ← Vq ∪ {v}
13: end for
14: end for
15: for each q ∈ Q do
16: µq ← Mean(Vq)
17: σq ← StandardDeviation(Vq)
18: Tstats ← Tstats ∪ {(q, µq, σq)} ▷ Compile final distributions
19: end for
20: return Tstats
21: end function
Algorithm 3 Ocular Verification and Multimodal Score-Level Fusion
1: Input: Frame variants F, Video variants V, Models M = Miris ∪ Mperi (Miris =
{Miris}), Condition sets Viris, Vperi, Periocular pairs D
orig
peri, Fusion weights W
2: Output: Unimodal & Fusion metrics Teval (EER, FMR, FNMR, d
′
, AUC)
3: Part 1: Unimodal Verification (Iris & Periocular)
4: function UnimodalVerification(F, V,M, Viris, Vperi)
5: Iiris ← {i | CheckQuality(Miris, f(OAC)
i
)} ▷ Indices of OAC-quality survivors
6: Teval ← ∅
7: for each modality X ∈ {iris, peri}, variant v ∈ VX, and model m ∈ MX do
8: Feval ← {f
(v)
i
: i ∈ Iiris} if X = iris else RandomSample(V
(v)
, k = 3)
9: SX,v,m ← ComputeScores(m, Feval)
10: Teval ← Teval ∪ {(X, v, m, ComputeStats(SX,v,m))}
11: end for
12: return Teval
13: end function
14: Part 2: Multimodal Score-Level Fusion
15: function ScoreLevelFusion(D
orig
peri, Miris, W)
16: Siris, Speri, Y ← ∅, ∅, ∅ ▷ Y represents ground truth labels (0/1)
17: for each pair (Ip, Ir, y, speri) ∈ D
orig
peri do ▷ Process 1. 8M exact pairs
18: siris ← 0. Ocular Verification for Virtual Reality 7
Algorithm 2 ISO Metric Quality Evaluation and Statistical Analysis
1: Input: Evaluated Frame Dataset F ∈ {Forig, FOAC , FUR, FUNIR}, BIQT engine EBIQT ,
Set of Target ISO Metrics Q
2: Output: Dataset Statistical Summary Tstats (containing µ and σ for all q ∈ Q)
3: function EvaluateQuality(F, EBIQT , Q)
4: Tstats ← ∅
5: for each q ∈ Q do
6: Vq ← ∅ ▷ Initialize empty sequence for metric values
7: end for
8: for each frame f ∈ F do
9: R ← RunBiqt(f, EBIQT ) ▷ Generate JSON quality report
10: for each q ∈ Q do
11: v ← ParseMetric(R, q) ▷ Extract scalar quality value
12: Vq ← Vq ∪ {v}
13: end for
14: end for
15: for each q ∈ Q do
16: µq ← Mean(Vq)
17: σq ← StandardDeviation(Vq)
18: Tstats ← Tstats ∪ {(q, µq, σq)} ▷ Compile final distributions
19: end for
20: return Tstats
21: end function
Algorithm 3 Ocular Verification and Multimodal Score-Level Fusion
1: Input: Frame variants F, Video variants V, Models M = Miris ∪ Mperi (Miris =
{Miris}), Condition sets Viris, Vperi, Periocular pairs D
orig
peri, Fusion weights W
2: Output: Unimodal & Fusion metrics Teval (EER, FMR, FNMR, d
′
, AUC)
3: Part 1: Unimodal Verification (Iris & Periocular)
4: function UnimodalVerification(F, V,M, Viris, Vperi)
5: Iiris ← {i | CheckQuality(Miris, f(OAC)
i
)} ▷ Indices of OAC-quality survivors
6: Teval ← ∅
7: for each modality X ∈ {iris, peri}, variant v ∈ VX, and model m ∈ MX do
8: Feval ← {f
(v)
i
: i ∈ Iiris} if X = iris else RandomSample(V
(v)
, k = 3)
9: SX,v,m ← ComputeScores(m, Feval)
10: Teval ← Teval ∪ {(X, v, m, ComputeStats(SX,v,m))}
11: end for
12: return Teval
13: end function
14: Part 2: Multimodal Score-Level Fusion
15: function ScoreLevelFusion(D
orig
peri, Miris, W)
16: Siris, Speri, Y ← ∅, ∅, ∅ ▷ Y represents ground truth labels (0/1)
17: for each pair (Ip, Ir, y, speri) ∈ D
orig
peri do ▷ Process 1. 8M exact pairs
18: siris ← 0.Where the paper says it
Where the paper says it
Where the paper says it
Performance Analysis
The results indicate that periocular recognition alone substantially outperforms unimodal iris verification across unconstrained VR feeds. The performance gap occurs due to the resilience of periocular structures to corneal reflections and off-axis distortions. p. 8 Therefore, the low observed mean score is due
to off-axis camera setup rather than evidence of poor image quality. Findings of Ocular Verification and Fusion Table 3 reports verification
performance on near-infrared periocular frames extracted from the VRBiom [10] Ocular Verification for Virtual Reality 7
Algorithm 2 ISO Metric Quality Evaluation and Statistical Analysis
1: Input: Evaluated Frame Dataset F ∈ {Forig, FOAC , FUR, FUNIR}, BIQT engine EBIQT ,
Set of Target ISO Metrics Q
2: Output: Dataset Statistical Summary Tstats (containing µ and σ for all q ∈ Q)
3: function EvaluateQuality(F, EBIQT , Q)
4: Tstats ← ∅
5: for each q ∈ Q do
6: Vq ← ∅ ▷ Initialize empty sequence for metric values
7: end for
8: for each frame f ∈ F do
9: R ← RunBiqt(f, EBIQT ) ▷ Generate JSON quality report
10: for each q ∈ Q do
11: v ← ParseMetric(R, q) ▷ Extract scalar quality value
12: Vq ← Vq ∪ {v}
13: end for
14: end for
15: for each q ∈ Q do
16: µq ← Mean(Vq)
17: σq ← StandardDeviation(Vq)
18: Tstats ← Tstats ∪ {(q, µq, σq)} ▷ Compile final distributions
19: end for
20: return Tstats
21: end function
Algorithm 3 Ocular Verification and Multimodal Score-Level Fusion
1: Input: Frame variants F, Video variants V, Models M = Miris ∪ Mperi (Miris =
{Miris}), Condition sets Viris, Vperi, Periocular pairs D
orig
peri, Fusion weights W
2: Output: Unimodal & Fusion metrics Teval (EER, FMR, FNMR, d
′
, AUC)
3: Part 1: Unimodal Verification (Iris & Periocular)
4: function UnimodalVerification(F, V,M, Viris, Vperi)
5: Iiris ← {i | CheckQuality(Miris, f(OAC)
i
)} ▷ Indices of OAC-quality survivors
6: Teval ← ∅
7: for each modality X ∈ {iris, peri}, variant v ∈ VX, and model m ∈ MX do
8: Feval ← {f
(v)
i
: i ∈ Iiris} if X = iris else RandomSample(V
(v)
, k = 3)
9: SX,v,m ← ComputeScores(m, Feval)
10: Teval ← Teval ∪ {(X, v, m, ComputeStats(SX,v,m))}
11: end for
12: return Teval
13: end function
14: Part 2: Multimodal Score-Level Fusion
15: function ScoreLevelFusion(D
orig
peri, Miris, W)
16: Siris, Speri, Y ← ∅, ∅, ∅ ▷ Y represents ground truth labels (0/1)
17: for each pair (Ip, Ir, y, speri) ∈ D
orig
peri do ▷ Process 1. 8M exact pairs
18: siris ← 0. For periocular recognition experiments, we randomly sampled 3 frames per video resulting in ∼110 frames per subject. 2 Metrics
We report the results in terms of (1) FNMR@FMR= 10−2
: False Non-Match
Rate (FNMR) at a specific False Match Rate (FMR), lower is better; (2) Equal
Error Rate (EER): denotes operating point when FMR and FNMR are equal,
lower is better; (3) Area Under the Curve (AUC): denotes the area under
the receiver operating point characteristics (ROC) curve, higher is better; (4)
d-prime (d’): measures the distinguishability between genuine and imposter
distributions in biometrics, typically, used in signal detection theory. Ocular Verification for Virtual Reality 7
Algorithm 2 ISO Metric Quality Evaluation and Statistical Analysis
1: Input: Evaluated Frame Dataset F ∈ {Forig, FOAC , FUR, FUNIR}, BIQT engine EBIQT ,
Set of Target ISO Metrics Q
2: Output: Dataset Statistical Summary Tstats (containing µ and σ for all q ∈ Q)
3: function EvaluateQuality(F, EBIQT , Q)
4: Tstats ← ∅
5: for each q ∈ Q do
6: Vq ← ∅ ▷ Initialize empty sequence for metric values
7: end for
8: for each frame f ∈ F do
9: R ← RunBiqt(f, EBIQT ) ▷ Generate JSON quality report
10: for each q ∈ Q do
11: v ← ParseMetric(R, q) ▷ Extract scalar quality value
12: Vq ← Vq ∪ {v}
13: end for
14: end for
15: for each q ∈ Q do
16: µq ← Mean(Vq)
17: σq ← StandardDeviation(Vq)
18: Tstats ← Tstats ∪ {(q, µq, σq)} ▷ Compile final distributions
19: end for
20: return Tstats
21: end function
Algorithm 3 Ocular Verification and Multimodal Score-Level Fusion
1: Input: Frame variants F, Video variants V, Models M = Miris ∪ Mperi (Miris =
{Miris}), Condition sets Viris, Vperi, Periocular pairs D
orig
peri, Fusion weights W
2: Output: Unimodal & Fusion metrics Teval (EER, FMR, FNMR, d
′
, AUC)
3: Part 1: Unimodal Verification (Iris & Periocular)
4: function UnimodalVerification(F, V,M, Viris, Vperi)
5: Iiris ← {i | CheckQuality(Miris, f(OAC)
i
)} ▷ Indices of OAC-quality survivors
6: Teval ← ∅
7: for each modality X ∈ {iris, peri}, variant v ∈ VX, and model m ∈ MX do
8: Feval ← {f
(v)
i
: i ∈ Iiris} if X = iris else RandomSample(V
(v)
, k = 3)
9: SX,v,m ← ComputeScores(m, Feval)
10: Teval ← Teval ∪ {(X, v, m, ComputeStats(SX,v,m))}
11: end for
12: return Teval
13: end function
14: Part 2: Multimodal Score-Level Fusion
15: function ScoreLevelFusion(D
orig
peri, Miris, W)
16: Siris, Speri, Y ← ∅, ∅, ∅ ▷ Y represents ground truth labels (0/1)
17: for each pair (Ip, Ir, y, speri) ∈ D
orig
peri do ▷ Process 1. 8M exact pairs
18: siris ← 0. Ocular Verification for Virtual Reality 7
Algorithm 2 ISO Metric Quality Evaluation and Statistical Analysis
1: Input: Evaluated Frame Dataset F ∈ {Forig, FOAC , FUR, FUNIR}, BIQT engine EBIQT ,
Set of Target ISO Metrics Q
2: Output: Dataset Statistical Summary Tstats (containing µ and σ for all q ∈ Q)
3: function EvaluateQuality(F, EBIQT , Q)
4: Tstats ← ∅
5: for each q ∈ Q do
6: Vq ← ∅ ▷ Initialize empty sequence for metric values
7: end for
8: for each frame f ∈ F do
9: R ← RunBiqt(f, EBIQT ) ▷ Generate JSON quality report
10: for each q ∈ Q do
11: v ← ParseMetric(R, q) ▷ Extract scalar quality value
12: Vq ← Vq ∪ {v}
13: end for
14: end for
15: for each q ∈ Q do
16: µq ← Mean(Vq)
17: σq ← StandardDeviation(Vq)
18: Tstats ← Tstats ∪ {(q, µq, σq)} ▷ Compile final distributions
19: end for
20: return Tstats
21: end function
Algorithm 3 Ocular Verification and Multimodal Score-Level Fusion
1: Input: Frame variants F, Video variants V, Models M = Miris ∪ Mperi (Miris =
{Miris}), Condition sets Viris, Vperi, Periocular pairs D
orig
peri, Fusion weights W
2: Output: Unimodal & Fusion metrics Teval (EER, FMR, FNMR, d
′
, AUC)
3: Part 1: Unimodal Verification (Iris & Periocular)
4: function UnimodalVerification(F, V,M, Viris, Vperi)
5: Iiris ← {i | CheckQuality(Miris, f(OAC)
i
)} ▷ Indices of OAC-quality survivors
6: Teval ← ∅
7: for each modality X ∈ {iris, peri}, variant v ∈ VX, and model m ∈ MX do
8: Feval ← {f
(v)
i
: i ∈ Iiris} if X = iris else RandomSample(V
(v)
, k = 3)
9: SX,v,m ← ComputeScores(m, Feval)
10: Teval ← Teval ∪ {(X, v, m, ComputeStats(SX,v,m))}
11: end for
12: return Teval
13: end function
14: Part 2: Multimodal Score-Level Fusion
15: function ScoreLevelFusion(D
orig
peri, Miris, W)
16: Siris, Speri, Y ← ∅, ∅, ∅ ▷ Y represents ground truth labels (0/1)
17: for each pair (Ip, Ir, y, speri) ∈ D
orig
peri do ▷ Process 1. 8M exact pairs
18: siris ← 0.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Future Directions
Future work will focus on extending their analysis to more datasets and designing an end-to-end image restoration and multimodal verification pipeline with revised quality evaluation. The authors also intuit that complex restoration networks can impair verification by over-smoothing fine iris patterns and may require domain adaptation. p. 3 We intuit that we
can leverage the underlying reflection removal mechanism on near-infrared data. Non-Uniform Illumination Restoration (UNIR-Net): Images captured
by VR headsets exhibit non-uniform illumination due to near-infrared illuminators and variations in lightning across the eye region, resulting in uneven brightness and reduced visibility of iris. Furthermore, the traditional ISO/IEC metric framework for margin adequacy and concentricity does
not work on VR-native imagery. Thus, we recommend that ISO/IEC framework
should accommodate diverse acquisition challenges present in VR devices and
revise thresholds or metrics for reliable validation of VR-based image quality. However, traditional iris recognition protocols may not be adequate in cases
of unconstrained acquisition, which is typical of VR-based data. In this
work, we examine three crucial aspects: (1) evaluating ISO/IEC 29794-6
iris quality metrics on VRBiom dataset and analyzing their limitations,
(2) addressing data-specific challenges such as off-axis gaze, non-uniform
illumination, and specular reflection using generative models, and (3)
performing unimodal (iris, periocular) recognition and multimodal scorelevel fusion (iris + periocular).Where the paper says it
Where the paper says it
Where the paper says it
--- Page 1 ---
Ocular Verification for Virtual Reality
Husanpreet Singh, Robert Tran, Ayushree Kharel, and Sudipta Banerjee
1 University of Wyoming, USA
2 San Diego State University, USA
8 Husanpreet Singh, Robert Tran, Ayushree Kharel, and Sudipta Banerjee 2 Related Work
--- Page 2 ---
Husanpreet Singh, Robert Tran, Ayushree Kharel, and Sudipta Banerjee 1 Iris Recognition
--- Page 3 ---
Husanpreet Singh, Robert Tran, Ayushree Kharel, and Sudipta Banerjee 2.1 Iris Recognition Classical iris recognition progressed from Daugman’s pioneering rubber sheet unwrapping and Gabor phase demodulation [5], to Wildes’ Hough transform boundary fitting [25], and Ma et al. p. 1 2 Related Work
2. 1 Iris Recognition
Classical iris recognition progressed from Daugman’s pioneering rubber sheet
unwrapping and Gabor phase demodulation [5], to Wildes’ Hough transform
boundary fitting [25], and Ma et al. 1 Iris Recognition
Classical iris recognition progressed from Daugman’s pioneering rubber sheet
unwrapping and Gabor phase demodulation [5], to Wildes’ Hough transform
boundary fitting [25], and Ma et al. ’s radial intensity variation tracking [12]
(reviewed comprehensively by Bowyer et al.Where the paper says it
Where the paper says it
--- Page 4 ---
Husanpreet Singh, Robert Tran, Ayushree Kharel, and Sudipta Banerjee 3 Proposed Approach Our proposed approach consists of three distinct processing modules.
--- Page 5 ---
Ocular Image Adjustments We address the off-axis perspectives, hardware-induced specular reflections, and uneven or non-uniform illumination in VR data using three operations.
--- Page 6 ---
Husanpreet Singh, Robert Tran, Ayushree Kharel, and Sudipta Banerjee 3.1 Ocular Image Adjustments Off-axis Correction (H8Net+DINOv3) Algorithm 1 details the pipeline that uses geometric off-angle correction (co) to mathematically estimate and mitigate the perspective distortion inherent to off-axis VR headset sensor arrays. p. 3 1 Ocular Image Adjustments
We address the off-axis perspectives, hardware-induced specular reflections, and
uneven or non-uniform illumination in VR data using three operations. Off-axis Correction (H8Net+DINOv3): Algorithm 1 details the pipeline
that uses geometric off-angle correction (co) to mathematically estimate and mitigate the perspective distortion inherent to off-axis VR headset sensor arrays.Where the paper says it
--- Page 7 ---
Husanpreet Singh, Robert Tran, Ayushree Kharel, and Sudipta Banerjee 3.3 Ocular Verification and Fusion We performed ocular verification firstly using unimodal schemes: (i) iris recognition using ArcIris, (ii) periocular recognition using MobileFaceNet. p. 4 We utilize the procedure
detailed in Algorithm 2 to systematically process our datasets, providing researchers with both the statistical spread of current VR imagery. 3 Ocular Verification and Fusion
We performed ocular verification firstly using unimodal schemes: (i) iris recognition using ArcIris, a deep learning-based model submitted to NIST IREX X
evaluation by UND, (ii) periocular recognition using lightweight mobilefacenetbased resnet architecture.Where the paper says it
--- Page 8 ---
Husanpreet Singh, Robert Tran, Ayushree Kharel, and Sudipta Banerjee 4.1 Dataset We utilize the publicly available VRBiom dataset introduced by Kotwal et al. p. 7 8 Husanpreet Singh, Robert Tran, Ayushree Kharel, and Sudipta Banerjee
4 Experiments and Analysis
4. 1 Dataset
We utilize the publicly available VRBiom dataset introduced by Kotwal et
al.Where the paper says it
--- Page 9 ---
Husanpreet Singh, Robert Tran, Ayushree Kharel, and Sudipta Banerjee 4.3 Implementation Details We used the MITRE BIQTIris framework and the internal implementation developed by University of Notre Dame (UND) with fine segmentation mask for ISO/IEC 29794-6 metric evaluation. p. 7 5 × (σ
2
gen + σ
2
imp)
(assuming similarity scores), higher is better. 3 Implementation Details
We used the MITRE BIQTIris framework2 and the internal implementation developed by University of Notre Dame (UND) with fine segmentation mask for
ISO/IEC 29794-6 metric evaluation.Where the paper says it
--- Page 10 ---
Husanpreet Singh, Robert Tran, Ayushree Kharel, and Sudipta Banerjee Table 1. Mathematical definitions of standard ISO/IEC 29794-6 quality metrics. p. 5 6 Husanpreet Singh, Robert Tran, Ayushree Kharel, and Sudipta Banerjee
Table 1. Mathematical definitions of standard ISO/IEC 29794-6 quality metrics.Where the paper says it
--- Page 11 ---
Ocular Verification for Virtual Reality
Table 3. Periocular Recognition Performance: Each cell reports Original / UnReflect / UNIR-Net processed images of VRBiom [10]
--- Page 12 ---
Husanpreet Singh, Robert Tran, Ayushree Kharel, and Sudipta Banerjee Table 4. Performance evaluation metrics for unimodal and multimodal fusion schemes on the VRBiom dataset
--- Page 13 ---
Ocular Verification for Virtual Reality
Table 4. Performance evaluation metrics for unimodal and multimodal fusion schemes on the VRBiom dataset
--- Page 14 ---
Husanpreet Singh, Robert Tran, Ayushree Kharel, and Sudipta Banerjee References
--- Page 15 ---
Ocular Verification for Virtual Reality
--- Page 16 ---
No references found in this section. (Note: The paper's pages are labelled '--- Page N ---' counting from 1 — the same N as that page's lines)
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Off-axis gaze correction using DINOv3 and H8Net can mathematically estimate and mitigate perspective distortion inherent to off-axis VR headset sensor arrays, restoring spatial proportions prior to polar unwrapping phase. This allows iris recognition systems to function more reliably when acquired from non-frontal angles.
-
Specularity removal via UnReflectAnything using a frozen vision transformer backbone and an inpainting module can reconstruct reflection free images by leveraging the underlying reflection removal mechanism on near-infrared data. This directly addresses the issue where
specular glare, and reflections introduced by the VR headset’s internal near-infrared (NIR) illuminators mask the fine-grained details of the iris pattern.
-
Non Uniform Illumination Restoration using UNIR-Net can mitigate non uniform illumination by combining illumination enhancement, attention mechanisms, visual refinement, and contrast correction to improve image quality and visibility. This helps counteract
uneven brightness and reduced visibility of iris
caused by near-infrared illuminators and lighting variations. -
Implement a multimodal score level fusion scheme that combines iris recognition scores with periocular recognition scores using a weighted formula like Fusion 25/75 (Peri Heavy) to achieve the best performance, yielding an EER of 0.33 and a decidability index d' of 0.99. This strategy boosts overall verification accuracy by approximately three times compared to unimodal iris recognition alone.
-
Develop a system that performs ocular verification using both iris recognition (ArcIris) and periocular recognition (MobileFaceNet) in parallel, followed by weighted score level fusion to enhance verification performance. This approach leverages the
resilience of periocular structures to corneal reflections and off-axis distortions
while compensating for theinherent degradations of unimodal iris recognition.
Abstract
Virtual reality (VR) headsets (e.g., Meta Quest, Apple Vision Pro) provide a seamless user experience due to their fast, frictionless interaction with the physical world in a simulated environment. User authentication relies on biometric cues such as iris in such headsets. However, traditional iris recognition protocols may not be adequate in cases of unconstrained acquisition, which is typical of VR-based data. In this work, we examine three crucial aspects: (1) evaluating ISO/IEC 29794-6 iris quality metrics on VRBiom dataset and analyzing their limitations, (2) addressing data-specific challenges such as off-axis gaze, non-uniform illumination, and specular reflection using generative models, and (3) performing unimodal (iris, periocular) recognition and multimodal score-level fusion (iris + periocular). We observe that some metrics (e.g., margin adequacy) fail on VR-acquired data; whereas, image adjustments primarily benefit periocular recognition, and multimodal fusion lowers EER by 11% over unimodal iris recognition performance. We will release the evaluation scripts upon acceptance for reproducibility.
Sources
- Biometrics in Extended Reality: A Review
- Lowering the Barrier to IREX Participation: Open-Source Algorithms, Toolkit, and Benchmarking for Iris Recognition
- VRBiom: A New Periocular Dataset for Biometric Applications of HMD
- GazeShift: Unsupervised Gaze Estimation and Dataset for VR
- DINOv3
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs