EveLoad: Event-Camera Eye Movement Dataset

• EveLoad is an event-camera based eye-movement dataset featuring graded cognitive workload levels with a controlled, grid-based fixation paradigm.
• The dataset uses a N-back guided fixation task with 2000 events per frame and five consecutive frames to capture rapid ocular dynamics at microsecond resolution.
• It provides both subject-specific and mixed split benchmarks, ensuring robust workload recognition performance while reducing gaze-location confounds.

Searching arXiv for the specified paper to ground the article in the source record. EveLoad is an event-camera-based eye-movement dataset and benchmark for cognitive workload recognition that was introduced to address two limitations in prior work: the predominance of frame-based eye trackers or RGB cameras, and the use of free-viewing, visual-search, tracking, or otherwise weakly constrained tasks in which models may exploit gaze-location distributions rather than workload-related ocular dynamics. The dataset is built around a spatially constrained, task-driven fixation paradigm recorded with an event camera, and the associated benchmark defines six workload levels derived from memory load and stimulus pacing. The originating study presents EveLoad as, to the authors’ knowledge, the first event-based eye-movement dataset with graded cognitive workload annotations, and reports strong performance for six-class workload recognition under both subject-specific and mixed random split evaluation (Lu et al., 23 Jun 2026).

1. Concept and motivation

EveLoad was created as a new event-camera-based eye-movement dataset and benchmark for cognitive workload recognition. Its central design premise is that workload recognition should be based on how the eyes move under cognitive load, rather than where participants look. Earlier workload datasets often permit substantial variation in gaze distributions across tasks or classes, which creates a confound: a classifier may learn spatial target statistics instead of cognitive workload cues. EveLoad addresses this by using a spatially constrained, task-driven fixation paradigm with event-camera sensing, so that workload-related ocular dynamics are observed under comparable gaze geometry (Lu et al., 23 Jun 2026).

The paper motivates the dataset for two main reasons. First, event cameras are presented as a particularly suitable sensing modality for eye movements because they offer microsecond-level temporal resolution, high dynamic range, low latency, and reduced motion blur. These properties are especially relevant for rapid ocular motion such as saccades and for subtle dynamics during fixation. Second, prior workload datasets are described as suboptimal for isolating workload-related eye behavior because many are collected under free-viewing or similarly weakly constrained conditions, allowing gaze-location distributions to vary substantially across tasks (Lu et al., 23 Jun 2026).

The application context is workload-aware rehabilitation, assistive interfaces, and XR systems. In these settings, cognitive workload monitoring is relevant because task difficulty, pacing, feedback, or rest scheduling may need to be adjusted according to the user’s cognitive state. EveLoad is positioned as a preliminary but important step toward such adaptive systems rather than as a clinical outcome study (Lu et al., 23 Jun 2026).

2. Dataset acquisition and sensing configuration

The dataset comprises recordings from 20 healthy adult participants, of whom 12 were female and 8 male. Total recording time is 7.5 hours, corresponding to about 22.5 minutes per participant, and the dataset contains 24,000 stimuli. Participants had normal or corrected-to-normal vision and were seated about 40 cm from a 1920 × 1080 monitor. Data collection was conducted under natural ambient lighting in order to better reflect realistic use conditions (Lu et al., 23 Jun 2026).

Eye movements were recorded using a DAVIS346 event camera positioned to capture the eye. The device records asynchronous brightness changes with microsecond temporal resolution. Stimulus presentation, keyboard responses, and event timestamps were synchronized through a unified logging system, enabling precise alignment between task events and the event stream. This synchronization is important because the benchmark is organized at the stimulus level, so each segment of event data must be associated with the corresponding task condition and behavioral response (Lu et al., 23 Jun 2026).

A concise summary of the acquisition setup is given below.

Component	Specification
Participants	20 healthy adult participants
Sex distribution	12 female, 8 male
Total recording time	7.5 hours
Per-participant duration	About 22.5 minutes
Total stimuli	24,000 stimuli
Viewing distance	About 40 cm
Display	1920 × 1080 monitor
Sensor	DAVIS346 event camera
Lighting	Natural ambient lighting

This configuration suggests that the dataset was designed to prioritize temporal fidelity and controlled behavioral conditions over multimodal breadth. A plausible implication is that EveLoad is especially suitable for studying fine-grained ocular dynamics under tightly specified task structure, while remaining narrower in population and scenario diversity than broader ecological datasets.

3. Controlled N-back-guided fixation paradigm

The defining design contribution of EveLoad is a controlled spatially constrained and task-driven N-back-guided fixation paradigm. Participants performed a continuous fixation task in which target positions were sampled from a predefined grid over an effective display region of about 1900 × 1000 pixels. The paper excludes 10 pixels on the left and right edges and 40 pixels on the top and bottom edges. Grid spacing was 100 × 100 pixels, yielding 190 fixed fixation points with roughly uniform spatial coverage (Lu et al., 23 Jun 2026).

Each fixation sequence contained 200 target locations, covered the 190 predefined fixation points, and constrained displacement between consecutive targets to 200–800 pixels. This was intended to ensure sufficient eye movement while avoiding repetitive local patterns and to reduce bias from any single target region. The resulting protocol maintains comparable spatial coverage across conditions, thereby reducing the chance that class discrimination depends primarily on fixation position (Lu et al., 23 Jun 2026).

The dataset defines six workload levels by combining memory load and stimulus pacing. Memory load is varied across no load, 0-back, and 1-back. Stimulus pacing is varied across slow and fast conditions. The six levels are listed explicitly in the study.

Level	Memory load	Stimulus	Decision rule	ISI
L1	none	green dot	fixate on stimulus	1500 ms
L2	none	green dot	fixate on stimulus	750 ms
L3	0-back	letter patch	match to letter H	1500 ms
L4	0-back	letter patch	match to letter H	750 ms
L5	1-back	letter patch	match to previous letter	1500 ms
L6	1-back	letter patch	match to previous letter	750 ms

Operationally, the six classes correspond to no-load slow, no-load fast, 0-back slow, 0-back fast, 1-back slow, and 1-back fast. In the no-load condition, participants fixate only on a green dot, providing a baseline with minimal cognitive demand. In the 0-back condition, they decide whether the current letter is the target letter H. In the 1-back condition, they decide whether the current letter matches the previous letter, thereby increasing working-memory demand. The slow and fast pacing conditions manipulate time pressure while keeping the cognitive task type fixed (Lu et al., 23 Jun 2026).

Stimulus details were also standardized. The no-load target was a green circular dot with 15 px diameter. The load conditions used a pixelated 100 × 100 patch containing a letter stimulus. Letters were drawn from H, I, O, X, and E. Participants responded with Enter if the condition was satisfied and Space otherwise (Lu et al., 23 Jun 2026).

4. Annotation structure and benchmark design

EveLoad uses a structured, reproducible labeling scheme centered on stimulus-level annotation. Each stimulus record includes fixation target coordinates $(x, y)$, the presented letter stimulus and target response, the participant’s response, stimulus start and end timestamps, and participant and trial IDs. This allows each stimulus segment of event data to be aligned with the corresponding fixation target, workload level label, and behavioral response (Lu et al., 23 Jun 2026).

The emphasis on stimulus-level labels is important because the benchmark supports official stimulus-level splits for evaluation. The paper states that EveLoad supports both subject-specific and mixed-subject evaluation at the stimulus level. In the subject-specific split, models are trained and tested separately for each participant using an 8:1:1 train/validation/test ratio. In the mixed-subject random split, stimuli from all participants are merged and randomly split, also with an 8:1:1 ratio, and samples from the same stimulus do not cross split boundaries (Lu et al., 23 Jun 2026).

This benchmark structure differentiates EveLoad from many prior workload datasets in four specific ways described by the paper: it uses event-based rather than frame-based sensing; it enforces controlled gaze distribution through a fixed grid with uniform spatial coverage; it provides finer workload granularity with six workload classes; and it supports benchmarking under subject-specific and mixed random splits. The authors also note that cross-dataset accuracy numbers are not directly comparable, which is a methodological caution against naïve leaderboard-style interpretation (Lu et al., 23 Jun 2026).

A common misconception in eye-movement-based workload recognition is that strong classification accuracy necessarily implies sensitivity to cognitive workload itself. EveLoad was explicitly designed to counter that concern by reducing gaze-location confounds. The paper does not claim complete elimination of all confounds, but its protocol aims to make workload-related ocular dynamics more salient than task-dependent spatial behavior (Lu et al., 23 Jun 2026).

5. Event representation and learning framework

The learning framework in EveLoad operates on event frames rather than raw events directly. Raw events are accumulated into frames using 2000 consecutive events per frame. Each frame is separated by polarity into positive and negative channels. Each sample contains 5 consecutive event frames, so the final input has 10 channels. The paper expresses the input tensor as

$X \in \mathbb{R}^{B \times C \times H \times W},$

where $B$ is batch size, $C$ the number of input channels, and $H \times W$ the event camera spatial resolution (Lu et al., 23 Jun 2026).

This stacking strategy encodes short-term temporal dynamics in the channel dimension while preserving spatial structure. A simple preprocessing step removes hot pixels, defined as pixels that continuously fire events at fixed positions and can dominate the representation. No extra spatial-temporal feature engineering was used beyond event-frame construction (Lu et al., 23 Jun 2026).

The paper compares three backbones: ResNet18, MobileNetV3, and MobileViT. The final model is a modified ResNet18 in which the first convolution is adapted to accept 10-channel input and initialized from ImageNet-pretrained weights. The backbone extracts high-level features, Global Average Pooling reduces feature maps to a compact vector, and the final fully connected layer maps the 512-dimensional feature vector to 6 workload logits (Lu et al., 23 Jun 2026).

The task is formulated as six-class classification with cross-entropy loss: $\mathcal{L}_{\text{cls} = -\frac{1}{N} \sum_{i=1}^{N} \log \frac{\exp(z_{i,y_i})} {\sum_{c=1}^{6} \exp(z_{i,c})},$ where $N$ is batch size, $\mathbf{z}_i \in \mathbb{R}^{6}$ are the logits for sample $i$, and $y_i \in \{0,\ldots,5\}$ is the class label (Lu et al., 23 Jun 2026).

Training used batch size 32, 20 epochs, learning rate $X \in \mathbb{R}^{B \times C \times H \times W},$0, and weight decay $X \in \mathbb{R}^{B \times C \times H \times W},$1. These details define the baseline benchmark rather than an exhaustive search over optimization regimes (Lu et al., 23 Jun 2026).

6. Experimental results and ablations

The paper reports both ablation and comparative results. For the event representation ablation, it varies the number of events per frame across 500, 1000, 2000, and 5000, and the number of consecutive frames $X \in \mathbb{R}^{B \times C \times H \times W},$2 across 1, 5, 10, and 15. The best result is obtained with 2000 events per frame and 5 frames, yielding 96.36% mean accuracy with 2.90% standard deviation; this setting is adopted as the final configuration (Lu et al., 23 Jun 2026).

Under this best input setting $X \in \mathbb{R}^{B \times C \times H \times W},$3, backbone comparison shows ResNet18 at 96.36%, MobileViT at 94.12%, and MobileNetV3 at 92.21%. The paper interprets this as evidence that the dataset benefits from a stronger convolutional representation than the lighter alternatives tested (Lu et al., 23 Jun 2026).

The main benchmark results are reported for two evaluation settings.

Evaluation setting	Accuracy	Macro F1
Subject-specific evaluation	96.36%	96.00%
Mixed random split evaluation	96.13%	95.84%

For subject-specific evaluation, the average accuracy is 96.36% and macro F1 is 96.00%, with consistently high performance across all 20 participants. For mixed random split evaluation, accuracy is 96.13% and macro F1 is 95.84%. The near equivalence between these two settings is interpreted in the paper as indicating that the learned workload structure is fairly stable across subjects when samples are controlled at the stimulus level (Lu et al., 23 Jun 2026).

The study further argues that the high performance under both evaluation settings supports the view that event-based ocular signals contain strong workload-discriminative information under the controlled fixation paradigm. This suggests that the model is learning workload-related spatiotemporal patterns rather than relying mainly on uncontrolled gaze-location biases. That inference is part of the paper’s interpretation rather than a direct causal proof (Lu et al., 23 Jun 2026).

7. Significance, limitations, and research scope

EveLoad is significant because it establishes a systematic benchmark for cognitive workload recognition from event-based eye movements under a controlled, spatially constrained, N-back-guided fixation paradigm. Within the scope defined by the study, it demonstrates that event cameras can capture workload-relevant ocular dynamics with very high accuracy and may therefore provide a useful sensing pathway for adaptive rehabilitation, assistive interfaces, XR/AR/VR systems, workload-aware closed-loop training, and real-time pacing and feedback adaptation (Lu et al., 23 Jun 2026).

Its distinction from prior datasets is not only sensor modality but also experimental control. Earlier cognitive workload datasets are described as typically relying on eye trackers, RGB cameras, VR headsets, or multimodal physiological sensors, often with 2–4 classes or binary workload labels and frequently under free-viewing or task-rich scenarios. EveLoad instead combines event-camera sensing, 6-class workload labeling, spatially constrained fixation, stimulus-level annotation, and subject-specific and mixed random split benchmarks (Lu et al., 23 Jun 2026).

The paper is explicit about the dataset’s limitations. EveLoad is not a clinical rehabilitation dataset; it is a preliminary exploration on healthy participants. Future validation is stated to be necessary in stroke, traumatic brain injury, and other neurological populations, and real-world deployment will require online integration into therapy systems. Accordingly, the current work functions as a benchmark and technical foundation rather than a clinical validation study (Lu et al., 23 Jun 2026).

A plausible implication is that EveLoad’s principal contribution lies in experimental design and sensing methodology rather than immediate translational readiness. For researchers, its value is methodological: it isolates workload-related ocular dynamics under controlled gaze geometry and provides a reproducible six-class event-based benchmark. For application-driven work, its limitations indicate that external validity across populations, tasks, and deployment settings remains an open problem (Lu et al., 23 Jun 2026).