The INI-30 Dataset : Event Camera for Eye Tracking

The Ini-30 dataset is collected with two event cameras mounted on a glass frame. Each DVXplorer sensor (640 × 480 pixels) is attached on the side of the frame. The power supply was provided via a 2 meter cable connected from the cameras to a computer, which provided enough freedom of movement. Differently from [2, 24], the participants were not instructed to follow a dot on a screen, but rather encouraged to look around to collect natural eye movements. As shown in Fig. 1, the event cameras were securely screwed on a 3D-printed case attached to the side of the glass frame. The data was annotated based on accumulated linearly decayed events by defining the pixel intensity as function of the linear accumulation of previous pixel intensity. Next we labeled the position of the pupil in the DVS’s array manually, using an assistive labeling tool. We discarded the first 20ms of events to ensure the eye was visible and annotations met the level of image-based annotators. The number of labels per recording was intentionally variable, spanning from 475 to 1’848 with a time per label ranging from 20.0 to 235.77 milliseconds depending on the overall duration of the sample. This setup allows for unconstrained head movements, enables to capture event data from eye movement in a ”in-the-wild” setting and allows the generation of a representative, unique, diverse and challenging dataset.NOTE : the annotations relates to the ellipse of the pupil on the imageIni-30数据集通过安装于玻璃框架的两台事件相机采集。每台DVXplorer传感器（分辨率640×480像素）固定于框架侧面。供电采用2米长线缆将相机连接至计算机，为设备提供充足的活动自由度。与相关研究[2,24]不同，本次实验未要求参与者追踪屏幕上的光点，而是鼓励其自由环顾以采集自然的眼球运动数据。如图1所示，事件相机通过3D打印外壳牢固安装在玻璃框架的侧面。数据标注基于累积线性衰减事件，将像素亮度定义为先前像素亮度的线性累积函数。随后，我们借助辅助标注工具，手动标注动态视觉传感器（Dynamic Vision Sensor, DVS）阵列中的瞳孔位置。为确保眼部可见且标注质量符合基于图像的标注标准，我们舍弃了前20毫秒的事件数据。每条录制样本的标注数量为可变值，范围从475至1848不等，单标签耗时从20.0毫秒至235.77毫秒不等，具体取决于样本的总时长。该装置支持无约束的头部运动，能够在“野外（in-the-wild）”场景下采集眼球运动的事件数据，可生成具备代表性、独特性、多样性且富有挑战性的数据集。注：本次标注对应图像中瞳孔的椭圆区域。