Ini-30: Event-Based Eye Tracking Dataset

• Ini-30 is an event-based eye-tracking dataset capturing asynchronous pupil data with precise pixel-level annotations under natural, mobile conditions.
• The dataset supports both dynamic and fixed event slicing strategies, enabling evaluation of neuromorphic models under variable temporal densities.
• Benchmark results indicate that Retina SNN achieves superior pupil localization accuracy and efficiency compared to competing models on this challenging dataset.

Ini-30 is an event-based eye-tracking dataset introduced in the context of neuromorphic pupil tracking and low-power deployment on spiking hardware. It was presented in "Retina: Low-Power Eye Tracking with Event Camera and Spiking Hardware" (Bonazzi et al., 2023) as a representative dataset built from pure event data captured by Dynamic Vision Sensor (DVS) cameras, with recordings from thirty volunteers under unconstrained, mobile conditions. The dataset is designed for pupil tracking and localization from asynchronous event streams rather than from conventional image frames, and its distinguishing features include glass-frame-mounted sensing, real-world movement variability, and pixel-level pupil annotations on the sensor.

1. Provenance and scope

Ini-30 was collected with two glass-mounted iniVation DVXplorer event-based cameras, one per eye, and comprises recordings from 30 volunteers (Bonazzi et al., 2023). The acquisition protocol did not require subjects to keep their head fixed or to follow a dot on a screen; participants were instead encouraged to move their heads and look around. Near Infrared illumination was used to enhance visibility of events around the pupil.

The dataset is intended for event-based pupil tracking in conditions closer to wearable use than laboratory eye-tracking setups. A central design choice is that the recordings are unconstrained and highly mobile, with no chinrest and no fixed-head requirement. This establishes Ini-30 as a benchmark for models that must operate under natural behavioral variation rather than under tightly controlled screen-gaze protocols.

A recurrent source of confusion is nomenclature. The Flickr popularity-prediction corpus used in "Predicting Popularity of Images Over 30 Days" is explicitly not called Ini-30; it is described instead as an extension of the SPID 2018 dataset for the ICIP 2020 Image Popularity Prediction Challenge (Dutta et al., 2021). In the event-based eye-tracking literature represented here, "Ini-30" refers to the pupil-tracking dataset introduced with Retina (Bonazzi et al., 2023).

2. Acquisition hardware and recording protocol

The sensing apparatus consisted of two iniVation DVXplorer DVS cameras with a native resolution of $640 \times 480$ pixels, mounted on a glass frame with a 3D-printed case, one camera per eye (Bonazzi et al., 2023). The cameras were powered through a 2-meter cable to a computer, permitting mobility during recording. Data from both cameras were streamed in real time.

Each event carries an $x, y$ pixel coordinate, a polarity, and a timestamp, yielding a 1-bit, truly asynchronous event stream. This representation is central to Ini-30: it is not frame-based and does not begin from synthetic frame reconstructions as the primary signal. Instead, raw DVS output is the native modality.

The recordings exhibit broad variability. Recording duration ranges from 14.64 s to 193.8 s per sequence, with 475 to 1,848 annotations and 4.8 million to 24.2 million events per recording. The interval between labels ranges from 20 ms to 235.77 ms. Event counts per sampled bin also vary substantially, with a minimum of 3, median 94, mean 175, standard deviation 299, and maximum 4,799 events. This variability is not incidental; it reflects the dataset’s emphasis on non-stationary, in-the-wild event statistics.

Property	Value
Cameras	2 iniVation DVXplorer DVS
Native resolution	$640 \times 480$ per eye
Participants	30 volunteers
Recording duration	14.64 s to 193.8 s
Labels per recording	475 to 1,848
Events per recording	4.8 million to 24.2 million
Label interval	20 ms to 235.77 ms
Events per bin	min 3, median 94, mean 175, max 4,799

These acquisition characteristics make Ini-30 particularly relevant for architectures that must handle irregular temporal density, variable motion, and the sensor statistics of wearable DVS systems.

3. Data representation and preprocessing pipeline

The native data format is the event tuple with spatial coordinate, timestamp, and polarity. For algorithmic compatibility, the recordings were cropped and center-aligned to a $512 \times 512$ square, then sum-pooled to $64 \times 64$, matching the resolution of the Speck neuromorphic SoC (Bonazzi et al., 2023). The resulting input preserves dual channels, one per polarity.

Event slicing supports two regimes. In the dynamic regime, each sample contains $N$ events irrespective of elapsed time. In the fixed regime, each sample contains all events within a fixed time window $\Delta t$. In both cases, each bin is centered on a label timestamp and includes events from the preceding span, forming an SNN input tensor of shape $C \times 64 \times 64$.

This preprocessing design has methodological consequences. Because Ini-30 can be partitioned either by event count or by elapsed time, it supports evaluation of algorithms under alternative notions of temporal normalization. A plausible implication is that the dataset is suitable not only for static benchmarking of accuracy, but also for studying how different event-slicing policies interact with neuromorphic inference latency, power, and temporal robustness.

4. Annotation protocol and ground truth

The ground truth is pixel-level annotation of pupil position in the cropped DVS array for each labeled time point (Bonazzi et al., 2023). Labels were assigned manually using a specialized tool. To facilitate annotation, the event stream was converted into synthetic intensity-like reconstructions using accumulated linearly decayed events. The first 20 ms of each recording were discarded to reduce warm-up noise.

For each annotation point, the dataset associates a timestamp, the $(x, y)$ pupil center on the DVS array, and the corresponding event slice leading up to that label. In the final SNN model discussed in the paper, outputs are represented through grid-based bounding boxes, while centroid position is used for comparison and evaluation.

The label semantics are notable. Ini-30 does not provide screen-coordinate gaze targets; it provides pupil localization directly on the sensor. This distinction matters because pupil-on-sensor annotation is closer to the immediate signal-processing problem solved by event-based wearable eye trackers, whereas screen-coordinate labels entangle sensor geometry, calibration, and projection. The paper explicitly contrasts Ini-30 with prior real event-based datasets that labeled gaze on screens rather than the pupil on the sensor (Bonazzi et al., 2023).

5. Benchmarking function and reported performance

The main benchmark task on Ini-30 is pupil tracking or localization on event slices (Bonazzi et al., 2023). Generalization is evaluated with leave-two-subjects-out cross validation. Reported metrics include centroid error in pixels, as well as power, latency, and energy when inference is deployed on neuromorphic hardware. Model size and MAC counts are also reported.

Within this benchmark, the Retina SNN is compared against the 3ET model. On Ini-30, 3ET achieves $4.48 \pm 1.94$ px centroid error, whereas Retina achieves $x, y$0 px. The paper also reports 418k parameters and 107M MACs for 3ET, versus 63k parameters and 3.03M MACs for Retina. On the Speck SoC, end-to-end power is reported between 2.89 and 4.80 mW, and latency between 5.57 and 8.01 ms, depending on whether dynamic or fixed windowing is used.

Model	Ini-30 error (px)	Params / MACs
3ET	$x, y$1	418k / 107M
Retina	$x, y$2	63k / 3.03M

These results position Ini-30 as more than a data repository: it functions as a hardware-conscious benchmark for neuromorphic eye-tracking pipelines. Because the paper reports both localization accuracy and edge-deployment characteristics, the dataset anchors evaluation across algorithmic and systems dimensions rather than accuracy alone.

6. Relation to prior datasets, uses, and access

The comparison reported in the Retina paper situates Ini-30 against earlier eye-tracking datasets characterized by lower variability, head-fixed acquisition, and label types based on screen coordinates rather than pupil position on the sensor (Bonazzi et al., 2023). In the comparison table, prior datasets identified as Angelopoulos20 / Zhao23 are listed with $x, y$3 sensor resolution, no glass-frame mounting, low variability, head-fixed movement constraints, and either mixed event/frame or synthetic inputs, whereas Ini-30 is described as $x, y$4, glass-frame mounted, high-variability, unconstrained, and pure-event.

This comparison is significant because it clarifies what Ini-30 is designed to measure. It is not optimized for conventional screen-based gaze estimation under rigid control; instead, it supports ultralow-power wearable eye tracking, neuromorphic and SNN research, and benchmarking under natural head and eye motion. The paper explicitly connects these properties to applications such as AR/VR glasses, assistive devices, and mobile diagnostics (Bonazzi et al., 2023).

A public link associated with the dataset and code is provided at https://github.com/pbonazzi/retina. The accompanying description indicates that this repository is the access point for the code and possibly for dataset access details. This suggests that reuse of Ini-30 is organized around the broader Retina benchmark and deployment stack rather than through a standalone archival release.

7. Conceptual significance

Ini-30 occupies a specific position in event-based vision research: it is a dataset for sensor-level pupil localization from real, asynchronous DVS streams collected under unconstrained motion (Bonazzi et al., 2023). Its combination of mobile acquisition, high event-count variability, manual pupil annotations, and explicit compatibility with low-power neuromorphic hardware gives it relevance to both algorithm design and systems evaluation.

Several misconceptions can be resolved directly from the published descriptions. First, Ini-30 is not a frame-based gaze dataset; its native modality is pure event data. Second, its labels are not screen coordinates but pupil positions in the DVS array. Third, it should not be conflated with unrelated datasets that also involve the number 30, such as the 30-day Flickr popularity dataset described in (Dutta et al., 2021). Taken together, these clarifications show that Ini-30 is best understood as a specialized benchmark for event-driven wearable eye tracking, particularly in settings where low power, low latency, and robustness to unconstrained motion are primary requirements.