EYE4ALL: Accessibility & OCT Vision Dataset

• EYE4ALL is a multifaceted research label combining an accessibility-oriented vision-language benchmark with an OCT dataset for anterior segment segmentation and eye biometry.
• The accessibility benchmark uses text-image-to-text generation refined by human annotators, focusing on spatial guidance, safety, and conciseness for blind/low-vision navigation.
• The OCT dataset comprises 6,621 volumes processed via deep learning segmentation and ray-tracing corrections, enabling personalized eye modeling and VR/AR calibration applications.

Searching arXiv for the named EYE4ALL papers and closely related supporting work. EYE4ALL is a name used for multiple research artifacts spanning accessibility-oriented vision-language evaluation, ophthalmic imaging, and, by extension in related literature, enabling technologies for assistive vision, eye tracking, and compact near-eye optics. In the most explicit recent usage, EYE4ALL denotes a blind/low-vision-centered text-image-to-text benchmark for evaluating large vision-LLM responses in navigation assistance (Kim et al., 1 Oct 2025). In another usage, it denotes an open-source whole-eye optical coherence tomography dataset built from a spectrally-multiplexed whole-eye OCT system for anterior segment segmentation, ocular biometry, and personalized eye modeling (Qian et al., 18 May 2026). Related work situates these EYE4ALL usages within a broader technical ecosystem that includes wearable assistive perception (Sethuraman et al., 2023), large-scale eye-image annotation for gaze and eye movement analysis (Fuhl et al., 2021), robust multi-view eye tracking (Arar et al., 2017), holographic cataract and intraocular lens simulation (Akyazi et al., 2023), and wide-field meta-optic eyepieces for near-eye systems (Wirth-Singh et al., 2024).

1. EYE4ALL as an accessibility benchmark for text-image-to-text evaluation

In the vision-language setting, EYE4ALL is a benchmark explicitly grounded in accessibility and blind/low-vision needs. It is built around text-image-to-text responses: an image plus a scene-relevant request are given to a large vision-LLM, and the resulting response is judged for how well it supports a blind or low-vision user navigating a real scene (Kim et al., 1 Oct 2025). The benchmark is intended to capture not merely image-text consistency, but whether a response is useful, safe, sufficiently detailed, concise, and free of hallucinations for blind/low-vision-oriented assistance.

The dataset is divided into two components. EYE4ALLPref is a pairwise preference dataset with chosen/rejected comparisons. EYE4ALLMulti is a pointwise multi-objective dataset with fine-grained human scores across seven dimensions: Direction Accuracy, Depth Accuracy, Safety, Sufficiency, Conciseness, Hallucination, and Overall Quality (Kim et al., 1 Oct 2025). The benchmark is constructed from pedestrian and navigation scenes drawn from the Sideguide and Sidewalk corpora, paired with requests described as plausible for blind/low-vision navigation contexts.

The generation pipeline begins with scene images, pairs them with a request, and then produces responses from Qwen2-VL, LLaVA-1.6, and InternLM-XComposer2-VL (InternLM-X2-VL). The models are prompted to answer from the perspective of blind or low-vision users. For each instance, one of the three model responses is selected randomly and then refined by GPT-4o mini using a blind/low-vision-oriented prompt (Kim et al., 1 Oct 2025). This makes EYE4ALL a curated, model-generated-and-refined benchmark rather than a corpus of purely human-authored responses.

The human annotation protocol is central to the benchmark’s design. The judgments were collected from 25 sighted human annotators, approved by the IRB. Each annotator completed 100 items online, taking about 2 to 3 hours. The guideline emphasized scene description, distance and direction to the target, obstacles to watch for, and step-by-step directions. The paper reports an average annotator agreement of 33.21 with a standard deviation of 17.70 (Kim et al., 1 Oct 2025). Non-hallucination dimensions are averaged from 2–3 human judgment scores on a 1–5 Likert scale, while Hallucination is binary, with 0 = presence and 1 = absence.

A distinctive feature of EYE4ALL is the separation of spatial guidance into direction and depth. The paper notes that Direction and Depth are separated because precise spatial guidance is especially important for blind or low-vision users (Kim et al., 1 Oct 2025). This reflects a benchmark design oriented toward embodied navigation rather than generic captioning quality.

2. EYE4ALL within multi-objective alignment modeling

EYE4ALL serves as the main accessibility-oriented evaluation target for MULTI-TAP (Multi-Objective Task-Aware Predictor), a plug-and-play architecture for single-objective and multi-objective image-text alignment prediction (Kim et al., 1 Oct 2025). In this context, EYE4ALL is not only a benchmark but also a training and evaluation substrate for human-aligned reward modeling in assistive scenarios.

MULTI-TAP is trained in two stages. In the first stage, a reward head predicts a single scalar alignment score from frozen or fine-tuned large vision-LLM hidden states using mean squared error:

$\min_\theta \sum^N_{i} (r_i - h_i)^2$

where $r_i$ is the predicted scalar score, $h_i$ is the human judgment score, and $\theta$ denotes the parameters of the large vision-LLM and reward head (Kim et al., 1 Oct 2025). In the second stage, a frozen large vision-LLM embedding $z_i$ is input to a ridge regression model to predict a vector of human-interpretable dimension scores:

$\hat y_i = W z_i + b$

with training objective

$$\min_{W,b}\ \sum_{i=1}^{N}\bigl\|y_i - W z_i - b\bigr\|_2^2 + \alpha\|W\|_F^2 .$$

The paper explicitly states that MULTI-TAP does not aggregate multi-objective outputs into a single score. Instead, the overall score comes from the first-stage reward head, while the dimension-specific scores come from the second-stage ridge regression head (Kim et al., 1 Oct 2025). This design is contrasted with models such as VisionREW-S, which aggregate multi-objective outputs with learned weights.

On EYE4ALLMulti, the strongest version of MULTI-TAP reaches about 52.08, whereas VisionREW-S achieves about 47.63 (Kim et al., 1 Oct 2025). The paper also emphasizes computational efficiency: VisionREW-S is reported to require about 51 days on a single RTX A6000 for the VisionREW setting, while MULTI-TAP completes training and inference in about 4 hours for a Qwen2-based variant and 11 hours for a LLaMA-3.2-based variant. A plausible implication is that EYE4ALL was designed not only as an evaluation instrument but also as a stress test for practical, interpretable, and efficient assistive-model assessment.

3. EYE4ALL as an open-source whole-eye OCT dataset

A second, unrelated but technically significant usage of the name EYE4ALL appears in ophthalmic imaging. Here, EYE4ALL denotes an open-source dataset derived from a spectrally-multiplexed whole-eye OCT (WEOCT) system, intended to fill a gap in public data for anterior segment segmentation, ocular biometry, and whole-eye reconstruction (Qian et al., 18 May 2026). The release contains 6,621 processed volumes from 276 unique participants, together with segmentation and calibrated 3D anterior point clouds.

The imaging system employs two synchronized swept sources: a 1310 nm anterior-segment channel with 60 nm bandwidth and 200 kHz sweep rate, and a 1060/1064 nm retinal channel with 102 nm bandwidth and 200 kHz sweep rate (Qian et al., 18 May 2026). The acquisition protocol uses

$$500 \text{ A-scans/B-scan}, \quad 250 \text{ B-scans/volume}$$

at a volumetric rate of about 1.6 Hz. The anterior channel covers roughly

$35 \text{ mm} \times 17.5 \text{ mm},$

while the retinal channel covers about

$36^\circ \times 18^\circ.$

The dataset includes three categories of released material: raw volumes, processed data, and manual annotations. Raw volumes comprise spectral OCT B-scan stacks and scan waveforms. Processed data include B-scans, segmentation masks, boundary pixels, full calibrated point clouds, and segmentation labels attached to the point clouds. Manual annotations are provided for a subset from 34 participants, with annotations available for each B-scan (Qian et al., 18 May 2026).

The processing pipeline has four stages: deep learning-based surface segmentation, 3D distortion correction, surface fitting, and ray-tracing refraction correction. The segmentation network is a DenseUNet trained on 51,398 annotated B-scans from 552 volumes from 30 subjects, with a validation set of 7,951 B-scans from 79 volumes from 8 subjects (Qian et al., 18 May 2026). The annotations label cornea/sclera and iris, with cornea and sclera grouped during annotation because their boundary is ambiguous in OCT.

Distortion calibration uses a custom dot-pattern target, a hexapod robot, multiple known Z positions, and per-pixel ray recovery, followed by 3rd-order 2D polynomial fits that map galvanometer voltages to physical 3D rays (Qian et al., 18 May 2026). After surface fitting, ray-tracing refraction correction accounts for the fact that OCT measures optical path length rather than true geometric depth. The pipeline uses fixed refractive indices

$r_i$0

and applies Snell’s law,

$r_i$1

at the relevant interfaces.

Validation is reported through both phantom and human studies. On a 12.5 mm radius ceramic sphere, the paper reports a median radius error of 14 μm, p95 radius error of 19 μm, median 3D center localization error of 30 μm, and p95 of 33 μm (Qian et al., 18 May 2026). In the phantom pupil-center experiment, the median error is 22 μm and p95 is 65 μm. The human study framework includes 300+ participants in the abstract framing, with 276 unique participants in the released QA-passed dataset.

4. Annotation, geometry, and eye-modeling context

The biomedical EYE4ALL release belongs to a broader trend toward geometry-rich eye datasets. A key comparator is TEyeD, which provides over 20 million real-world eye images collected with seven head-mounted eye trackers, including devices integrated into VR and AR systems (Fuhl et al., 2021). TEyeD includes 2D and 3D landmarks, semantic segmentations, 3D eyeball annotation, gaze vectors, and eye movement types.

TEyeD’s annotations span pupil, iris, and eyelids in both 2D and 3D, plus gaze and movement labels such as fixations, saccades, smooth pursuits, blinks, and invalid/error cases (Fuhl et al., 2021). The dataset uses a semi-supervised Multiple Annotation Maturation (MAM) workflow, beginning with 20,000 manually annotated images, and iteratively refining labels with a ResNet-50 landmark regressor and a residual U-Net segmentation model. Images are flagged for manual review when the Jaccard agreement between derived segmentations falls below 0.9, using

$r_i$2

A notable point in TEyeD is the distinction between the pupil’s image-plane appearance and its physical 3D position. Because of corneal refraction and steep camera angles, the pupil often appears displaced relative to the iris center; accordingly, the dataset adjusts 3D landmarks and 3D segmentation to the iris center rather than directly using the observed pupil appearance (Fuhl et al., 2021). This geometry-aware treatment complements the biometric goals of the OCT-based EYE4ALL release. This suggests that the two datasets address adjacent but distinct layers of eye modeling: TEyeD is image-centric and behavior-centric, whereas the OCT EYE4ALL is anatomy-centric and biometry-centric.

The relevance to personalized eye modeling is explicit in the OCT work, which states applications including visual/optical axis measurement, eye tracking model calibration, VR/AR/MR eye modeling, and personalized gaze estimation (Qian et al., 18 May 2026). In that sense, EYE4ALL as whole-eye OCT infrastructure contributes anatomical priors and calibrated geometry that can underpin future gaze-estimation and XR calibration pipelines.

5. Assistive-wearable interpretations and related perceptual systems

Although not itself named EYE4ALL, MagicEye provides a closely related assistive-wearable model that helps situate accessibility-oriented interpretations of the term. MagicEye is an intelligent wearable eyeglass-based system for visually impaired users that integrates real-time object detection, facial recognition, currency recognition, GPS-based navigation, and proximity-based obstacle alerts into one device (Sethuraman et al., 2023).

Its object detection core is based on YOLOv5, with a CSPNet backbone, PANet, and a YOLO detection head that outputs 20×20, 40×40, and 80×80 feature maps (Sethuraman et al., 2023). The detector is trained on 35 classes selected from Open Images Dataset V4, using images resized to 416 × 416, batch size of 32, and SGD with momentum. The paper reports mAP = 68.2% on 1661 held-out samples, using a training set of over 13,200 samples.

The paper also reports distinct subsystem metrics. Currency identification reaches 99.75% accuracy, 99.86% F1-score, 100% recall, and 99.72% precision. Facial recognition reaches 94.5% accuracy, 94.21% F1-score, 89.60% recall, and 99.34% precision (Sethuraman et al., 2023). Face recognition uses an ensemble of FaceNet and VGG-Face, with MTCNN for face detection and cosine similarity

$r_i$3

for matching.

MagicEye is architecturally relevant because it frames assistive vision as a multi-modal always-available wearable. Its pipeline begins when the user activates the device or the proximity sensor triggers it, followed by image capture, AI processing, object detection, audio feedback, optional face and currency identification, GPS-based route guidance, and immediate obstacle warnings (Sethuraman et al., 2023). This suggests one plausible practical interpretation of EYE4ALL as a broader research direction: a convergence of multimodal sensing, model-based perception, and accessibility-focused output, even when the name itself is used more narrowly in individual papers.

6. Eye tracking, near-eye optics, and perceptual display infrastructure

The EYE4ALL-related literature also includes enabling hardware and algorithms for robust eye sensing and near-eye image delivery. In gaze tracking, Robust Real-Time Multi-View Eye Tracking introduces a multi-camera framework that simultaneously acquires multiple eye appearances and fuses the resulting gaze outputs through adaptive weighting (Arar et al., 2017). The prototype uses three synchronized PointGrey Flea3 monochrome cameras, each with 1280×1024 resolution, 8 mm lenses, and 850 nm near-infrared LEDs, and runs at 30 fps.

The system computes up to $r_i$4 gaze outputs per frame for $r_i$5 cameras and fuses them as

$r_i$6

with head-pose-based or gazing-behavior-based weights (Arar et al., 2017). In the reported experiments, the multi-view system achieves about 1 degree accuracy under challenging scenarios and nearly 100% availability, including improvements for users with glasses.

At a lower-level ocular-feature scale, Custom Video-Oculography Device and Its Application to Fourth Purkinje Image Detection during Saccades presents a full-resolution MJPEG video-oculography platform supporting offline reanalysis of pupil, First Purkinje image (P1), and Fourth Purkinje image (P4) detection (Abdulin et al., 2019). The system saves every frame at full resolution, supports recordings up to 500 fps, and uses a staged blob-selection procedure for P4 detection inside a pupil-centered area of interest. Candidate P4 blobs must have area between 5 and 30 pixels, maximum brightness above an adaptive threshold, and are selected by proximity to the pupil center. The paper shows that a P1 − P4 signal qualitatively resembles historical dual-Purkinje-image tracking during saccades (Abdulin et al., 2019).

On the display side, Wide Field of View Large Aperture Meta-Doublet Eyepiece demonstrates a wide-field meta-optic eyepiece relevant to compact near-eye systems (Wirth-Singh et al., 2024). The paper reports a meta-doublet eyepiece with greater than 60° field of view and an entrance aperture of 2.1 cm, designed at 633 nm. For the 2 cm meta-doublet, reported specifications include 21.0 mm entrance aperture, 60° full FoV, 5.4 mm pupil diameter, 15.0 mm eye relief, 15.17 mm effective focal length, 0.18 numerical aperture, and 35.7 mm total track length (Wirth-Singh et al., 2024). A comparable refractive triplet is listed at 43.0 mm total track length and 45° apparent FoV.

The meta-doublet consists of periodic arrays of square silicon nitride pillars on quartz, with 750 nm tall pillars, widths from 80 to 270 nm, 350 nm pitch, SiN $r_i$7, and quartz substrate $r_i$8 (Wirth-Singh et al., 2024). Phase delays are computed with rigorous coupled-wave analysis (RCWA) and optimized in Zemax OpticStudio. The first metasurface acts as an entrance aperture and corrective plate, while the second provides most of the focusing power. For near-eye accessibility devices, a plausible implication is that such optics could eventually support thinner and lighter monocular or binocular perceptual-assistance displays.

A related perceptual-simulation direction appears in Artificial Eye Model and Holographic Display Based IOL Simulator, which combines an artificial pseudophakic eye with a holographic simulator for preoperative IOL counseling (Akyazi et al., 2023). The benchtop system uses a phase-only spatial light modulator (LCoS SLM) with 1920 × 1080 resolution, 8 µm pixel pitch, and 60 fps phase modulation, together with He-Ne laser, wavelength 632.9 nm, and LED, wavelength 530 nm illumination (Akyazi et al., 2023). It evaluates a monofocal IOL: 14.5 D and a bifocal IOL: 15.5 + 3.25 D, showing halos, glare, contrast loss, and coherence-dependent interference fringes under laser illumination.

7. Conceptual scope, limitations, and recurring themes

Across the cited literature, EYE4ALL does not denote a single unified research program. Rather, it names at least two specific resources in different subfields: an accessibility-centered text-image-to-text benchmark (Kim et al., 1 Oct 2025) and an OCT-based whole-eye imaging dataset (Qian et al., 18 May 2026). The surrounding literature shows recurring concerns with robustness, human-centered evaluation, geometry-rich annotation, and deployable sensing or display hardware.

In the accessibility benchmark usage, the central concern is whether large vision-LLM outputs are safe, sufficient, spatially accurate, concise, and non-hallucinatory for blind or low-vision navigation assistance (Kim et al., 1 Oct 2025). In the OCT usage, the emphasis is on anatomically accurate 3D reconstruction, segmentation, and biometry, enabled by spectrally multiplexed acquisition and physics-based correction (Qian et al., 18 May 2026). These are distinct technical agendas, but both are organized around making eye- and vision-related systems more practically useful.

Several limitations are explicit in the source papers. The benchmark paper states that even strong models remain limited on challenging multi-objective tasks such as EYE4ALL, and that hallucination-free, human-preferred blind/low-vision-oriented generation remains open (Kim et al., 1 Oct 2025). The OCT dataset paper notes substantial scan rejection from motion and poor retinal visibility, with about 45% of collected volumes acceptable after QA (Qian et al., 18 May 2026). MagicEye leaves long-term battery life, full embedded deployment constraints, and failure modes under poor lighting or crowded scenes largely unexamined (Sethuraman et al., 2023). TEyeD includes invalid and no-eye frames and depends partly on geometrically derived annotations (Fuhl et al., 2021). The meta-optic eyepiece remains monochromatic at 633 nm, which constrains it mainly to monochrome near-eye display and night-vision contexts (Wirth-Singh et al., 2024).

Taken together, the literature indicates that EYE4ALL functions less as a singular technical object than as a recurring label at the intersection of accessibility, ophthalmic measurement, and eye-centered computational imaging. One branch evaluates whether multimodal AI can assist blind or low-vision users safely and effectively in real scenes (Kim et al., 1 Oct 2025). Another provides calibrated anatomical data for ocular reconstruction and biometry (Qian et al., 18 May 2026). Surrounding work on wearable assistance (Sethuraman et al., 2023), eye-image annotation (Fuhl et al., 2021), multi-view tracking (Arar et al., 2017), Purkinje-based oculography (Abdulin et al., 2019), holographic ocular simulation (Akyazi et al., 2023), and meta-optic eyepieces (Wirth-Singh et al., 2024) defines the broader technical context in which EYE4ALL-related systems may develop.