GazeMining: A Dataset of Video and Interaction Recordings on Dynamic Web Pages. Labels of Visual Change, Segmentation of Videos into Stimulus Shots, and Discovery of Visual Stimuli.
收藏资源简介：
<strong>Recording setup</strong><br> Recordings have been taken place on 12th March 2019. Gaze data has been recorded with a Tobii 4C eye tracker with Pro license at 90 Hz. Resolution of the viewport was set to 1024x768. The display had a size of 24 inches and a resolution of 1680x1050 pixels. We polled the DOM tree every 50 milliseconds for fixed elements. We recorded the Web browsing of four participants, who followed the protocol as stored under "Dataset_visual_change/Instructions.doc". <strong>Description of the dataset</strong><br> The dataset consists of following three subsets. <em>1. Dataset_visual_change</em><br> The recordings of each participant p1-p4 on twelve Web sites are in the corresponding directories. For each Web site, there are nine to eleven files: <site>.json: datacast <site>.webm: video recording <site>.features.csv: computer-vision features per observation <site>.features_meta.csv: meta information about features <site>.labels-l<X>.csv: labels of observations <site>_meta.csv: meta information about recording <site>_scroll_cache.csv: cache of estimated scrolling <site>_scroll_cache_map.csv: mapping of observations to scroll cache entries <site>_times.csv: timestamps of frames in the video recording <site>_layer_pixels.csv: first row is the pixel count of root layer, second row is pixel count of all fixed elements <em>2. Dataset_stimuli</em><br> Stimulus shots and visual stimuli computed with the framework. Value-based, edge-based, signal-based, and SIFT-based features have been used. The labels of the first participant's session had been used to train a random forest classifier with 100 trees for visual change classification, using the named features. The discovery has been performed on each Web site from the dataset and<br> the results are placed in the respective directories. Inside each directory, there is one directory for the detected shots and one for the discovered stimuli. In the shots directory, there is one overview as <participant>_<site>.csv file. For each shot, there are four further files: <participant>_<site>_<shot>.png: stitched frame of the stimulus shot <participant>_<site>_<shot>-blind.csv: frames from animations that are not contributing to the stitched frame <participant>_<site>_<shot>-gaze.csv: gaze data (in stitched frame space) <participant>_<site>_<shot>-mouse.csv: mouse data (in stitched frame space) The shots have been merged to stimuli, which are placed in the stimuli directory. The stimuli are grouped per layer (scrollable, fixed elements, etc.) and meta information is available in <layer_index>-<xpath>-meta.csv files. Furthermore, there are directories per layer, storing the discovered stimuli. Each discovered visual stimulus is represented by four files: <stimulus_id>.png: stitched frame of the visual stimulus <stimulus_id>-gaze.csv: gaze data (in stitched frame space) <stimulus_id>-mouse.csv: mouse data (in stitched frame space) <stimulus_id>-shots.csv: contained stimulus shots <em>3. Dataset_evaluation</em><br> We have performed two evaluations of the visual stimuli discovery. One computational estimating the quality of stimuli. One case-study of an expert's task. There are two respective directories with the annotation data. <strong>Changelog</strong><br> [1.0.2] Add counts of layer pixels per participant.<br> [1.0.1] Change to CC0 license.<br> [1.0.1] Add labels of third annotator "l3".<br> [1.0.0] Initial release.
<strong>录制设置</strong> 本次录制于2019年3月12日完成。眼动数据通过搭载专业版授权的Tobii 4C眼动仪采集，采样率为90 Hz。视口分辨率设为1024×768。显示器尺寸为24英寸，物理分辨率为1680×1050像素。我们每50毫秒轮询一次文档对象模型（Document Object Model，DOM）树以检测固定元素。本次采集了4名参与者的网页浏览行为数据，所有参与者均遵循存储于"Dataset_visual_change/Instructions.doc"中的实验流程。 <strong>数据集说明</strong> 本数据集包含以下三个子数据集： <em>1. 视觉变化子数据集（Dataset_visual_change）</em> 4名参与者（p1至p4）在12个网站上的录制数据分别存储于对应目录中。每个网站对应9至11个数据文件，具体如下： <site>.json：数据广播（datacast）文件 <site>.webm：视频录制文件 <site>.features.csv：单观测点的计算机视觉特征文件 <site>.features_meta.csv：特征元信息文件 <site>.labels-l<X>.csv：观测点标注文件，其中X为标注者编号 <site>_meta.csv：录制任务元信息文件 <site>_scroll_cache.csv：预估滚动行为缓存文件 <site>_scroll_cache_map.csv：观测点与滚动缓存条目映射文件 <site>_times.csv：视频录制帧时间戳文件 <site>_layer_pixels.csv：首行为根图层像素数，次行为所有固定元素像素总数 <em>2. 视觉刺激子数据集（Dataset_stimuli）</em> 本部分包含通过本框架计算得到的刺激镜头与视觉刺激数据，采用了基于价值、基于边缘、基于信号以及基于尺度不变特征变换（Scale-Invariant Feature Transform，SIFT）的多种特征。研究使用上述特征，以第一名参与者会话的标注数据训练了包含100棵决策树的随机森林分类器，用于视觉变化分类任务。本研究在数据集的所有网站上执行了视觉刺激发现任务，结果存储于对应目录中。 每个网站目录下包含两个子目录：分别存储检测到的镜头与发现的视觉刺激。在镜头目录中，存在一个总览文件「<participant>_<site>.csv」。每个镜头对应4个附属文件： <participant>_<site>_<shot>.png：刺激镜头拼接帧图像 <participant>_<site>_<shot>-blind.csv：未参与拼接的动画帧数据 <participant>_<site>_<shot>-gaze.csv：拼接帧坐标系下的眼动数据 <participant>_<site>_<shot>-mouse.csv：拼接帧坐标系下的鼠标操作数据 所有镜头将被合并为视觉刺激，存储于刺激子目录中。视觉刺激将按图层（可滚动图层、固定元素图层等）分组，元信息存储于「<layer_index>-<xpath>-meta.csv」文件中。此外，每个图层对应独立子目录以存储发现的视觉刺激，每个视觉刺激包含4个文件： <stimulus_id>.png：视觉刺激拼接帧图像 <stimulus_id>-gaze.csv：拼接帧坐标系下的眼动数据 <stimulus_id>-mouse.csv：拼接帧坐标系下的鼠标操作数据 <stimulus_id>-shots.csv：该刺激包含的镜头列表 <em>3. 评估子数据集（Dataset_evaluation）</em> 本数据集包含两项针对视觉刺激发现任务的评估：一项为量化评估视觉刺激质量的计算实验，另一项为专家任务案例研究。两类评估的标注数据分别存储于对应的两个子目录中。 <strong>更新日志</strong> [1.0.2] 新增各参与者的图层像素统计数据 [1.0.1] 采用CC0开源许可协议 [1.0.1] 新增第三名标注者「l3」的标注数据 [1.0.0] 首次正式发布
