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

TL;DR

ImmerIris introduces a large-scale, off-axis iris dataset captured with VR headsets to address immersive recognition challenges such as perspective distortion, intra-subject variation, and quality degradation. It provides eight evaluation protocols that systematically probe isolated and combined distortions, enabling robust benchmarking of traditional and deep learning methods. The authors propose NormFree, a normalization-free end-to-end approach using a cropped iris region and ArcFace loss, which consistently outperforms normalization-based SOTAs under immersive conditions. The results reveal that state-of-the-art methods trained on traditional, controlled data do not generalize well to immersive scenarios, highlighting the need for dataset-driven benchmarks and end-to-end solutions in this domain.

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 564 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.

Figures (6)

• Figure 1: Comparison of application scenarios. Each sample group is from the same person. (a) Traditional iris recognition acquires on-axis and controlled images with dedicated devices, with samples being highly invariant. (b) Immersive iris recognition collects images using consumer HMDs, yielding off-axis and unconstrained samples that exhibit distortion, variation, and degradation.
• Figure 2: Performance of SOTAs and our normalization-free approach on CASIA-Iris-V4 casia-iris-v4 and 4 increasingly challenging ImmerIris protocols. Lower FRR is better. SOTAs perform well on the traditional setup yet drop sharply under immersive conditions, whereas our approach consistently outperforms them.
• Figure 3: Data acquisition setup. (a) Screen interface of the VR headset, where red squares numbered 1-9 mark gaze points for sequential fixation. Live camera previews assist proper wearing. A full-screen white panel gradually increases in brightness to simulate illumination changes. (b) Actual scene of data acquisition.
• Figure 4: Sample ocular images. (a) Images that fail annotation are obviously flawed and therefore removed. (b) Normal samples.
• Figure 5: Data annotation. (a) Quality scores across 6 dimensions. Samples with quality scores below the thresholds (vertical lines) are considered degraded, and the rest as normal. (b) Examples from each quality-degradation category.