[翻译+注解] DEAP dataset：a dataset for emotion analysis using eeg, physiological and video signals

DEAP dataset
An emotion analysis dataset using EEG, physiological and video signals

1. abstract, abstract

We present a multimodal dataset for analyzing human emotional states. Electroencephalogram (EEG) and peripheral physiological signals were recorded while 32 participants watched 40 one-minute-long music video excerpts. Participants rated each video on levels of arousal, pleasantness, like/dislike, dominance, and familiarity. For 22 of the 32 participants, forehead facial videos were also recorded. A novel stimulus selection method was used, leveraging emotion tag retrieval from the last.fm website, video highlight detection and online assessment tools.

This dataset is publicly available, and we encourage other researchers to use it to test their own affective state estimation methods. This dataset was first introduced in the paper: " DEAP: A Database for Emotion Analysis using Physiological Signals (PDF) ", S. Koelstra, C. Muehl, M. Soleymani, J.-S. Lee, A. Yazdani , T. Ebrahimi, T. Pun, A. Nijholt, I. Patras, EEE Transactions on Affective Computing, vol. 3, no. 1, pp. 18-31, 2012

---

What is an electroencephalogram (EEG)?

Electroencephalography (EEG) is a non-invasive neurophysiological technique used to measure the electrical activity of the brain. It works by placing a series of electrodes on the scalp to record the brain's electrical signals, which reflect the activity of the brain's neurons. EEG is widely used in clinical medicine, neuroscience research, and brain-computer interfaces.

In an EEG, the electrical signals captured by the electrodes are presented in the form of a graph, usually a series of waveforms. These waveforms include oscillations of different frequencies such as:

• $Delta$ wave (Delta wave): The frequency range is usually between 0.5 and 4 Hertz (Hz), mainly related to deep sleep.
• $Theta$ waves: Frequency ranges from 4 to 8 Hz, commonly associated with sleep, relaxation, and brain activity in children.
• $Alpha$ wave (Alpha wave): frequency range is 8 to 13 Hz, usually occurs in the resting state, especially when the eyes are closed.
• $Beta$ waves: Frequency ranges from 13 to 30 Hz, usually associated with states of alertness, attention, and thought activity.
• $Gamma$ waves: Frequency range is above 30 Hz, related to advanced cognitive functions and information processing.

EEG is used in medicine to diagnose and monitor a variety of neurological conditions, such as epilepsy, sleep disorders, and brain injuries. In addition, it is used to study the basic functions of the brain, cognitive processes, emotional regulation and the development of brain-computer interface technology. EEG is an important tool that helps us understand brain activity patterns and health.

---

What frequency waves are related to emotions?

There is a certain correlation between brain waves of different frequencies and different emotional and cognitive states, although this correlation is not absolute because brain activity is complex. Here are some relationships between common emotional and cognitive states and brainwave frequencies:

1. $Alpha$ wave (Alpha wave):
   • Relax: $An\; increase\; in\; alpha$ waves is often associated with a state of relaxation, such as when you close your eyes, sit down and rest.
   • Introspection and Meditation: $Alpha$ waves are also associated with states of introspection, meditation, and deep concentration.
2. $\beta$ wave (Beta wave)：
   • Alertness and concentration: higher $Beta$ wave activity typically accompanies brain activity during alertness, concentration, and the performance of cognitive tasks.
   • Anxiety: Too much $Beta$ wave activity is sometimes associated with feelings of anxiety, but it's not the only factor.
3. $Theta$ wave：
   • Relaxation and creative thinking: θ θIncreased$theta\; wave\; activity\; may\; be\; associated\; with\; relaxation,\; creative\; thinking,\; and\; fantasy.$
   • Deep Meditation: In a state of deep meditation, θ θIncreased$theta\; wave\; activity.$
4. $Delta$ wave (Delta wave):
   • Deep sleep: $Delta$ waves occur primarily during deep sleep stages and are therefore associated with rest and recovery.
     
     It is important to note that emotional and cognitive states are affected by many factors, of which brain waves are only one. Furthermore, different emotional and cognitive states often involve multiple brain regions working together, not just fluctuations in specific frequencies.
     
     So while the frequency of brain waves can provide clues about brain activity, a full understanding of emotional and cognitive states requires consideration of other neurophysiological, psychological and environmental factors.

---

What is last.fm website about?

Last.fm is a music social networking site and internet radio platform. It allows users to create personal music profiles, record the music they listen to, interact with other music lovers, and discover new music. Last.fm also provides a personalized music recommendation service based on a user's listening history, as well as music event and concert information. With Last.fm, users can track their musical tastes, build music social circles, and discover content related to their musical interests.

In the field of research and data analysis, Last.fm's music tags and user behavior data are often used to build music recommendation systems, analyze music trends, and study users' music preferences and behaviors. Therefore, Last.fm's music tags and data can be used to create sentiment-labeled music datasets for sentiment analysis and research.

2. dataset summary, dataset overview

The DEAP dataset consists of two parts:

[ Part 1 ] Ratings from online self-assessments, where each one-minute clip from the music video was rated by 14 to 16 volunteers based on activation, pleasure, and dominance.

• Arousal : Measures whether the music video arouses excitement in the audience.
• Valence : Used to indicate whether the emotion conveyed by the music video is positive or pleasant.
• Dominance : reflects the degree of control the music video has over the audience. .

These ratings can be used to study how music and video content affects viewers' emotional and emotional experiences.

[ Part 2 ] Participant ratings, physiological recordings, and facial videos from an experiment in which 32 volunteers watched 40 subsets of the music videos mentioned above. Electroencephalogram (EEG) and physiological signals were recorded, and each participant also rated the videos according to the above criteria. For 22 participants, frontal facial videos were also recorded.

• Participant ratings : Similar to [Part 1], viewers rated the video's activation, pleasure, and dominance to understand their emotional response to the video content.
• Physiological recording : This includes recording of electroencephalogram (EEG) and other physiological signals. EEG data can be used to analyze brain activity to understand the cognitive and emotional responses of viewers while watching videos.
• Facial videos : For 22 participants, videos of their frontal faces were also recorded. This can be used to analyze the relationship between facial expressions and emotions, and how the viewer's emotional experience is reflected in facial expressions while watching the video.

For a more detailed explanation of the dataset collection and its contents, see [1] .

The 32 volunteers in the second part were different from the volunteers in the first part. In the first part, 14 to 16 independent volunteers watched different clips of the music video and rated the clips for emotion (activation, pleasure, and dominance). These volunteers were used to construct the first part of the data.

The 32 volunteers in the second part were another group of volunteers who watched the same or similar subset of music videos, but they were different individuals, excluding the volunteers in the first part. These Part II volunteers were also rated for emotion (activation, pleasure, and dominance), and physiological signals and facial videos were additionally recorded. This design was to study the relationship between emotion ratings and physiological/facial responses and differences in emotional experience in a different group of volunteers.

3. file listing, file list

The following files are available (detailed description of each file below):

file name	Format	part	content
Online_ratings	xls, csv, ods spreadsheets	Online self-assessment	All individual ratings in the online self-assessment
Video_list	xls, csv, ods spreadsheets	two parts	Name/YouTube link of the music video used in the online self-assessment and experiment, and individual rating statistics from the online self-assessment
Participant_ratings	xls, csv, ods spreadsheets	experiment	All ratings given by participants to the video during the experiment
Participant_questionnaire	xls, csv, ods spreadsheets	experiment	Questionnaires that participants answered before the experiment
Face_video	Zip file	experiment	Video recordings of foreheads and faces of participants 1-22 in the experiment
Data_original	Zip file	experiment	Unprocessed physiological data recordings from experiments, stored in BioSemi .bdf format
Data_preprocessed	Zip files for Python and Matlab	experiment	The physiological data recordings in the experiment were preprocessed (downsampling, EOG removal, filtering, segmentation, etc.) and stored in Matlab and Python (numpy) formats

4. file details, file details

4.1 online_ratings, online ratings

This file contains all individual video ratings collected during the online self-assessment. The file is available in Open-Office Calc ( online_ratings.ods), Microsoft Excel ( online_ratings.xls), and comma-separated values ​​( online_ratings.csv) formats.

Ratings were collected using an online self-assessment tool as described in [1] . Participants rated ① arousal (arousal) , ② valence (valence) , and ③ dominance (dominance) using discrete 9-point scales using SAM mannequins . In addition, participants rated felt emotions using an emotion wheel (see [2] ).

---

What are SAM mannequins?

" SAM " stands for "Self-Assessment Manikin," which is a tool for subjective affective assessment, often used to study and measure people's affective and emotional states. SAM manikins are a psychological measurement tool that allow subjects to graphically represent their emotional experiences .

SAM manikins typically consist of a set of human models with different expressions that represent different dimensions of emotional states. Typically, SAM manikins have three dimensions:

1. Arousal : This dimension represents the level of activity or excitement of the emotional state. In SAM manikins, subjects are asked to select a character model whose posture or expression best reflects their current level of activity. Typically, the persona on the left represents low arousal (e.g., calm, calm), while the persona on the right represents high arousal (e.g., excited, excited).
2. Valence : This dimension represents the positive and negative nature of the emotional state. In SAM manikins, subjects are asked to select a character model whose posture or expression best reflects whether their current emotion is positive or negative. Typically, the persona on the left represents negative emotions, while the persona on the right represents positive emotions.
3. Dominance : This dimension represents the degree of initiative or control of an individual in an emotional state. In SAM manikins, subjects are asked to select a character model whose posture or expression best reflects whether they currently feel they have control or dominate the situation. Typically, the upper persona represents high dominance, while the lower persona represents low dominance.

By selecting appropriate character models on SAM manikins, subjects can express their emotional states graphically, which makes emotional assessment more objective and standardized. SAM manikins are often used in experiments and investigations in fields such as affective psychology, emotion research, and human-computer interaction.

What is a discrete 9-point scale?

The discrete 9-point scale is a scoring method commonly used in psychology and emotion research. It is usually used to assess people's subjective feelings about a certain emotion or emotional attribute, such as pleasure, arousal, emotional intensity, etc. This scale divides the rating into 9 different levels, each level represents a different degree or feeling. Typically, these levels are distributed between two extremes, such as from very unpleasant to very pleasant, or from very calm to very excited. Participants were asked to select a level to express their subjective feelings or evaluations.

Discrete 9-point scales are typically used to obtain a discrete, limited number of subjective ratings rather than a continuous, unlimited number of ratings. This discrete scale can be easier to analyze and interpret because it limits the rating to several discrete choices rather than a continuous scale. In emotion research, pleasantness, arousal, and dominance are often subjectively assessed using this discrete 9-point scale.

What are arousal, valence and dominance?

In affective psychology, arousal, valence and dominance are the three key dimensions that describe emotions and emotions. They are often used to quantitatively measure and describe an individual's affective state and emotional experience.

1. Arousal: Arousal describes the level of physiological arousal or activity of an emotion or emotion. Higher arousal indicates that the individual is in a more excited, alert, or agitated state, while lower arousal indicates that the individual is in a calmer or calmer state. Arousal is often related to physiological indicators such as heart rate, electrodermal activity, etc.
2. Valence: Valence describes the positive and negative nature of an emotion or emotion. Higher pleasantness means that the emotion is more positive, pleasant, and positive, while lower pleasantness means that the emotion is more negative, unpleasant, and negative. Pleasure is often used to differentiate between positive emotions (such as happiness) and negative emotions (such as sadness) .
3. Dominance: Dominance describes an individual's sense of initiative or control in a specific emotional state. Higher levels of dominance indicate that the individual feels more dominant and in control, while lower levels of dominance indicate that the individual feels more dominated and less controlled. Dominance is usually used to describe the dominance of individual behaviors and emotions in emotional states.

"Dominance" is a dimension in emotional psychology that describes emotions and emotions. It may be relatively uncommon and difficult to understand. Let’s explain further:
Dominance describes an individual’s sense of agency or control in a certain emotional state. In other words, it measures whether an individual feels in control and dominance during a certain affective or emotional state, or whether they feel dominated by others or situations.

For example, if someone feels very confident, determined, and in control in a certain emotional state, that emotional state might be described as high in dominance. This means that the individual feels in this emotional state that he or she has the ability to dominate and control the situation.

On the other hand, if someone feels uncomfortable, helpless, and influenced by others in an emotional state, then that emotional state may be described as low on the dominance scale. This means that the individual feels less in control in this emotional state and is more easily influenced by external factors or others .

Dominance is part of the affective dimension and is often used together with arousal (the activeness of the emotion) and pleasantness (the positive or negative nature of the emotion) to more fully describe an individual's emotional experience. The understanding of dominance may vary across cultures, contexts, and fields of study, but it remains valuable when studying emotions and emotions because it provides insight into the impact of affective states on individual behavior and experiences.

These dimensions are often used to study emotions and emotions in order to more accurately understand and describe an individual's emotional experience. They can also be used to assess the impact of culture, products, or media content on individuals' emotions.

What is the emotion wheel?

The Emotion Wheel is a graphic tool used to describe and understand the range of emotions . It is usually presented in the form of a circular chart, with different emotional states or emotion categories arranged in different parts of the circle. The purpose of the emotion wheel is to help people identify, differentiate and express various emotions in order to better understand and communicate emotional experiences.





The emotion wheel is usually divided into different sections, with each section representing an emotion or category of emotion. These sections are often labeled with words or phrases that describe the nature of the emotion. The design of an emotion wheel can vary based on specific research or application needs, but typically includes a range of basic emotion states such as:

1. joy
2. anger
3. fear
4. sad
5. surprise
6. disgust

The Emotion Wheel can also further subdivide these basic emotional states to provide a more detailed description of the emotional experience. For example, joy can be divided into subcategories such as happiness, contentment, excitement, etc. This segmentation helps capture and express emotions more accurately.

The emotion wheel is often used in fields such as psychological research, emotion analysis, emotion recognition, psychotherapy, and emotional intelligence to help people better understand and manage emotions. It can also be used as a tool in emotion recognition systems to help computers recognize and understand human emotions.

The table in the file has one row for each individual score and contains the following columns:

List	describe
Online_id	Column corresponding to the same video ID in the video_list file
Valence	Pleasure rating (integer between 1 and 9)
Arousal	Arousal score (integer between 1 and 9)
Dominance	Dominance score (integer between 1 and 9)
Wheel_slice	Selected Sections on the Emotion Wheel For some participants, the emotion wheel ratings were not recorded correctly. In these cases, the value of Wheel_slice is 0. Otherwise, the mapping of emotions to integers on the emotion wheel is as follows: 1. Pride 2. Excitement 3. Joy 4. Contentment 5. Comfort 6. Hope 7. Interest 8. Surprise 9. Sadness 10. Fear 11. Shame 12. Guilt 13 . Jealousy 14. Disgust 15. Contempt 16. Anger
Wheel_strength	The intensity of the choice on the emotion wheel (integer from 0 to 4 → weak to strong)

4.2 video_list, video list

This file lists all the videos used in the online self-assessments and experiments, presented in a table format. This file is available in Open-Office Calc ( video_list.ods), Microsoft Excel ( video_list.xls), and comma-separated values ​​( video_list.csv) formats.

Each row of the table represents a video and contains the following columns:

List	describe
Online_id	Unique identifier used in online self-assessment
Experiment_id	If this video is selected for use in an experiment, lists the unique identifiers that will be used in the experiment. Blank if not selected
Lastfm_tag	If this video was selected via the last.fm sentiment tag, the sentiment tags are listed. Otherwise blank
Artist	the artist who recorded the song
Title	song title
Youtube_link	Download the original YouTube link of the video. Please note that due to copyright restrictions we are unable to provide the videos we use and the links may have been removed or not available in your country
Highlight_start	The time in seconds at which the extracted one-minute highlights started as determined by MCA analysis. For some videos, highlights are manually overwritten (for example, when a certain part of the song is particularly well-known)
Num_ratings	The number of volunteers who rated this video in the online self-assessment
VAQ_Estimate	The experimenter selected the pleasure/arousal quadrant of this video. For each quadrant, 15 videos were selected by last.fm and 15 were selected manually. These quadrants include: 1. High arousal, high pleasure 2. Low arousal, high pleasure 3. Low arousal, low pleasure 4. High arousal, low pleasure
VAQ_Online	Pleasure/arousal quadrant determined by average ratings from online self-assessment volunteers. Please note that these may differ from the estimated quadrants
AVG_x, STD_x, Q1_x, Q2_x, Q3_x	The mean, standard deviation and first, second and third quartiles of online self-assessment of pleasure/arousal/dominance by volunteers

4.3 participant_ratings, participant ratings

这个文件包含了在实验期间收集的所有参与者对视频的评分。该文件以 Open-Office Calc（participant_ratings.ods）、Microsoft Excel（participant_ratings.xls）和逗号分隔值（participant_ratings.csv）格式提供。

start_time 值是由演示软件记录的。愉悦度、唤醒度、支配度和喜好度是在每次试验后立即使用标准鼠标在连续的 9 点量表上评定的。愉悦度、唤醒度和支配度的评分使用 SAM 人体模型可视化呈现。对于喜好度（即您有多喜欢这个视频？），使用了拇指向上和拇指向下的图标。熟悉度是在实验结束后使用 5 点整数量表进行评定的（从“以前从未听过”到“定期听”）。遗憾的是，参与者 2、15 和 23 的熟悉度评分遗失了。

5 点整数量表是什么

5 点整数量表是一种常用于心理学、市场调查和社会科学研究中的评分方式。它通常用于评估人们对某个属性、产品、情感或观点的主观感受或态度。这种量表将评分分为 5 个离散的级别，每个级别代表着不同的态度或感受。通常，这些级别从极端的否定到极端的肯定排列，中间还包括一个中性选项。参与者需要选择一个级别，以表示他们对特定对象或主题的态度或评价。



5 点整数量表通常用于获取有限数量的离散主观评分，而不是连续的、无限数量的评分。这种量表简单明了，容易理解，适用于各种调查和研究领域。在市场调查中，常常使用这种量表来评估产品满意度、广告效果、品牌偏好等。在心理学研究中，它也可以用于评估情感、态度、焦虑程度等。

该文件中的表格每行对应一个参与者对视频的评分，包含以下列：

列名	列内容
Participant_id	参与者的唯一标识号（1 - 32）
Trial	试验编号（即呈现顺序）
Experiment_id	与 video_list 文件中相同列的视频编号
Start_time	试验视频播放的起始时间，单位为微秒（相对于实验开始时间）
Valence	愉悦度评分（浮点数，介于 1 和 9 之间）
Arousal	唤醒度评分（浮点数，介于 1 和 9 之间）
Dominance	支配度评分（浮点数，介于 1 和 9 之间）
Liking	喜好度评分（浮点数，介于 1 和 9 之间）
Familiarity	熟悉度评分（整数，介于 1 和 5 之间）。如果缺失，则为空白

4.4 participant_questionnaire，参与者问卷

该文件包含了实验前参与者填写的问卷答案。该文件提供了Open-Office Calc（participant_questionnaire.ods）、Microsoft Excel（participant_questionnaire.xls）和逗号分隔值（participant_questionnaire.csv）格式。

问卷中的大多数问题都是多项选择题，基本可以理解。不幸的是，参与者 26 没有填写问卷。此问卷还包括对同意书上的问题的回答（数据是否可以用于研究，您的影像是否可以发布）。

4.5 face_video.zip，正脸视频

Face_video.zip 包含了实验中记录的前 22 位参与者的正脸视频，已分割成各个试验。在 zip 文件中，sXX/sXX_trial_YY.avi 对应于主题 XX 的试验 YY 的视频。

对于参与者 3、5、11 和 14，由于技术问题（即录像带用完了），最后几次试验中的一个或多个试验缺失。请注意，这些视频按照呈现顺序排列，因此试验编号与 video_list 文件中的 Experiment_id 列不对应。试验编号与 Experiment_ids 之间的映射可以在 participant_ratings 文件中找到。

视频是从放置在屏幕后面的三脚架上以 DV PAL 格式使用 SONY DCR-HC27E 摄像机录制的。然后，将视频根据试验进行分割，并使用 h264 编解码器转码为 50 fps 的去隔行视频。转码是使用 mencoder 软件执行的，命令如下：

mencoder sXX.dv -ss trialYY_start_second -endpos 59.05 -nosound -of avi -ovc x264
  -fps 50 -vf yadif=1:1,hqdn3d -x264encopts bitrate=50:subq=5:8x8dct:frameref=2:bframes=3 
  -noskip -ofps 50 -o sXX_trialYY.avi

上面的代码是使用 mencoder 软件对录制的 DV PAL 格式视频文件 sXX.dv 进行处理和转码，以创建一个 AVI 视频文件 sXX_trialYY.avi，其中 XX 是主题编号，YY 是试次编号。

以下是代码的各个部分的解释：

• mencoder sXX.dv: 这部分指定了要处理的源视频文件，sXX.dv 是源文件名。
• -ss trialYY_start_second: 这部分指定了从源视频中的哪个时间点开始处理。trialYY_start_second 是试次 YY 的起始时间点（以秒为单位）。这样可以选择从源视频的特定时间点开始处理。
• -endpos 59.05: 这部分指定了要处理的视频持续时间。在这里，视频被裁剪为约 59.05 秒长。
• -nosound: 这部分告诉 mencoder 不包括音频。
• -of avi: 这部分指定输出文件的格式为 AVI。
• -ovc x264: 这部分指定了视频编解码器，使用 x264 编解码器进行视频压缩。
• -fps 50: 这部分指定输出视频的帧速率为 50 帧每秒。
• -vf yadif=1:1,hqdn3d: 这部分应用了视频滤镜。yadif 用于去除隔行扫描效果，hqdn3d 用于进行去噪处理。
• -x264encopts bitrate=50:subq=5:8x8dct:frameref=2:bframes=3: 这部分指定了 x264 编解码器的选项，包括比特率、子像素质量、8x8DCT 变换等。
• -noskip: 这部分禁止跳帧，确保所有帧都被编码。
• -ofps 50: 这部分指定输出视频的帧速率为 50 帧每秒。
• -o sXX_trialYY.avi: 这部分指定输出文件的名称，sXX_trialYY.avi 是输出文件的名称。

总之，这个代码片段的目的是将源视频文件进行处理，裁剪成指定持续时间的视频，并应用滤镜和编解码器，最终生成一个 AVI 格式的视频文件。

视频的同步精度大约为 1/25 秒（除了人为错误）。同步是通过在实验前后同时显示红色屏幕，并与发送到脑电图记录计算机的标记同步来实现的。然后，手动标记了该屏幕的起始帧在视频记录中的位置。然后，从脑电图记录中的试验开始标记计算各个试次的开始时间。

4.6 data_original.zip，原始数据

data_original.zip 中包含了原始的数据记录。共有 32 个 .bdf 文件（由 Actiview 录制软件生成的 BioSemi 数据格式），每个文件都有 48 个以 512Hz 采样的通道（32 个 EEG 通道，12 个外周通道，3 个未使用的通道和 1 个状态通道）。.bdf 文件可以由各种软件工具包读取，包括 Matlab 的 EEGLAB 和 BIOSIG 工具包。

数据是在两个不同的地点记录的。参与者 1-22 在 Twente 记录，参与者 23-32 在 Geneva 记录。由于硬件的不同版本，格式存在一些细微差异。首先，两个地点的 EEG 通道顺序不同。其次，每个地点的 GSR 测量数据格式也不同。

Twente 和 Geneva 是什么意思？

Twente 和 Geneva 是两个不同的地理位置或地点：

1. Twente：Twente 是荷兰东部的一个地区，也是荷兰的一个行政区域。在上下文中，Twente 可能指的是在该地区或该地区附近进行的实验或数据记录。
2. Geneva：Geneva 是瑞士的一座城市，位于瑞士法语区域，是国际外交和国际组织的重要中心。在上下文中，Geneva 可能指的是在该城市或该城市附近进行的实验或数据记录。

下表列出了两个地点的 EEG 通道名称（根据 10/20 系统）以及可用于将一种顺序转换为另一种顺序的索引：

通道编号	Twente通道名	Geneva通道名	Geneva > Twente	Twente > Geneva
1	Fp1	Fp1	1	1
2	AF3	AF3	2	2
3	F7	F3	4	4
4	F3	F7	3	3
5	FC1	FC5	6	6
6	FC5	FC1	5	5
7	T7	C3	8	8
8	C3	T7	7	7
9	CP1	CP5	10	10
10	CP5	CP1	9	9
11	P7	P3	12	12
12	P3	P7	11	11
13	Pz	PO3	16	14
14	PO3	O1	13	15
15	O1	Oz	14	16
16	Oz	Pz	15	13
17	O2	Fp2	32	30
18	PO4	AF4	31	29
19	P4	Fz	29	31
20	P8	F4	30	27
21	CP6	F8	27	28
22	CP2	FC6	28	25
23	C4	FC2	25	26
24	T8	Cz	26	32
25	FC6	C4	22	23
26	FC2	T8	23	24
27	F4	CP6	20	21
28	F8	CP2	21	22
29	AF4	P4	18	19
30	Fp2	P8	17	20
31	Fz	PO4	19	18
32	Cz	O2	24	17

这个表格列出了两个地点（Twente 和 Geneva）的 EEG 通道名称及其相应的编号。Geneva 到 Twente 和 Twente 到 Geneva 的映射也在表格中提供。这些通道名称和编号是根据国际 10/20 系统制定的，用于标识头皮上的不同脑电通道。

剩余的通道编号在两个地点都是相同的。然而，请注意两个地点的皮肤电导（GSR）测量采用不同的单位。Twente 的 GSR 测量单位是纳西门子（nano-Siemens），而 Geneva 的 GSR 测量单位是欧姆（Ohm）。转换关系如下：

$$\mathrm{GS}{\mathrm{R}}_{\mathrm{Geneva}}=10^9/GS{\mathrm{R}}_{\mathrm{Twente}}$$

下表解释了剩余通道的含义：

通道编号	通道名称	通道内容
33	EXG1	hEOG1（位于左眼左侧）
34	EXG2	hEOG2（位于右眼右侧）
35	EXG3	vEOG1（位于右眼上方）
36	EXG4	vEOG4（位于右眼下方）
37	EXG5	zEMG1（颧弓肌主要肌肉，距离嘴角左侧约1厘米）
38	EXG6	zEMG2（颧弓肌主要肌肉，距离zEMG1约1厘米）
39	EXG7	tEMG1（斜方肌，左肩胛骨）
40	EXG8	tEMG2（斜方肌，距离tEMG1约1厘米下方）
41	GSR1	皮肤电导反应，左手中指和无名指
42	GSR2	未使用
43	Erg1	未使用
44	Erg2	未使用
45	Resp	呼吸带
46	Plet	充血计，左拇指
47	Temp	温度，左小指
48	Status	状态通道，包含标记

状态通道包含从刺激呈现计算机发送的标记，指示试验的开始和结束。以下状态标记被使用：

状态码	事件持续时间	事件描述
1（第一次出现）	N/A	实验开始（参与者按键开始）
1（第二次出现）	120,000 毫秒	基线记录开始
1（进一步出现）	N/A	评分屏幕开始
2	1,000 毫秒	视频同步屏幕 （第一次试验之前，休息前后，最后一次试验后）
3	5,000 毫秒	试验开始前的固定屏幕
4	60,000 毫秒	音乐视频播放开始
5	3,000 毫秒	音乐视频播放后的固定屏幕
7	N/A	实验结束

4.7 data_preprocessed_matlab.zip and data_preprocessed_python.zip

这些文件包含了一个经过降采样（至 128Hz）、预处理和分段处理的数据版本，以 Matlab（data_preprocessed_matlab.zip）和 Python/Numpy（data_preprocessed_python.zip）格式提供。这个数据版本非常适合那些希望快速测试分类或回归技术而不必首先处理所有数据的人。每个 zip 文件包含 32 个 .dat（Python）或 .mat（Matlab）文件，每个参与者一个。以下是加载 Python 数据文件的示例代码：

import cPickle
x = cPickle.load(open('s01.dat', 'rb'))

每个参与者文件包含两个数组：

数组名称	数组形状	数组内容
data	$40\times 40\times 8064$	视频/试验 × 通道 × 数据
labels	$40\times 4$	视频/试验 × 标签 (愉悦度, 唤醒, 支配度, 喜好)

DEAP 的数据包括两个主要的数组：data 和 labels，其中：

1. data 数组：
   • 数组形状：$40\times 40\times 8064$
   • 数组内容：视视频/试验 × 通道 × 数据

这个数组的维度解释如下：

• 第一个维度（$40$）：代表了数据库中的不同视频或试验。换句话说，DEAP数据库包含了 $40$ 个独立的视频或试验的数据集。
• 第二个维度（$40$）：代表了每个视频或试验中采集的生理信号通道的数量。这意味着每个视频或试验包括 $40$ 个不同的生理信号通道，例如心率、皮肤电导、脑电图等。
• 第三个维度（$8064$）：代表了每个通道中采集的数据点数。这个维度表示了每个通道的时间序列数据的长度，每个通道中包含 $8064$ 个数据点。

---

2. labels 数组：
   • 数组形状：$40\times 4$
   • 数组内容：视频/试验 × 标签（愉悦度、唤醒度、支配度、喜好）

这个数组的维度解释如下：

• 第一个维度（$40$）：同样代表了数据库中的不同视频或试验。每一行对应于数据库中的一个视频或试验。
• 第二个维度（$4$）：代表了情感标签的种类，包括愉悦度、唤醒度、支配度和喜好。每个视频或试验都有与之关联的这四个情感标签的数值。

总之，DEAP 数据库中的 data 数组包含了 $40$ 个视频或试验的生理信号数据，每个视频/试验包括 $40$ 个不同的通道，每个通道中有 $8064$ 个数据点。而 labels 数组包含了这 $40$ 个视频/试验对应的愉悦度、唤醒度、支配度和喜好这四个情感标签的数值。这些数据可以用于情感分析和情感识别的研究。

视频按照 Experiment_id 的顺序排列，而不是按照演示顺序排列。这意味着每个参与者的第一个视频都是相同的。以下表格显示了通道布局和执行的预处理步骤：

通道编号	通道内容	预处理
1	Fp1
2	AF3
3	F3
4	F7
5	FC5
6	FC1
7	C3
8	T7
9	CP5
10	CP1
11	P3
12	P7
13	PO3
14	O1	1. 数据被降采样到128Hz
15	Oz	2. EOG（眼电图）伪迹被移除，方法如[1]所述
16	Pz	3. 应用了4.0-45.0Hz的带通频率滤波器
17	Fp2	4. 数据被平均到了共同的参考电极
18	AF4	5. EEG通道被重新排序，以使它们都按照上面提到的Geneva顺序排列
19	Fz	6. 数据被分割成60秒的试验，并去除了3秒的预试验基线
20	F4	7. 试验被重新排序，从演示顺序改为视频（Experiment_id）顺序
21	F8
22	FC6
23	FC2
24	Cz
25	C4
26	T8
27	CP6
28	CP2
29	P4
30	P8
31	PO4
32	O2
----	-------------------------------------------	-------------------------------------------
33	hEOG（水平眼电图，hEOG1 - hEOG2）
34	vEOG（垂直眼电图，vEOG1 - vEOG2）
35	zEMG（颧肌电图，zEMG1 - zEMG2）	1. 数据被降采样到128Hz
36	tEMG（斜方肌电图，tEMG1 - tEMG2）	2. 数据被分割成60秒的试验，并去除了3秒的试验前基线
37	GSR（从Twente转换为Geneva格式的皮肤电导数据（欧姆））	3. 试验被重新排序，从演示顺序改为视频（Experiment_id）顺序
38	呼吸带
39	测容量计
40	温度

5. dataset access，数据集访问

为了获取数据集并下载此页面上的文件，请访问 DEAP 数据集请求服务器，以获取用户名和密码。您需要使用它们来访问以下文件。请注意，该数据集的许可仅限于学术研究。请注意，只有在学术 / 研究机构拥有永久职位的人员（用户）才能申请。用户必须使用其所在机构的电子邮件地址申请，并提供链接到其所在机构的网页，以验证其在机构的职位和电子邮件地址。

5.1 metadata，元数据

包含在四个电子表格中的所有元数据（在线评分、视频列表、参与者评分和参与者问卷）。提供以下格式：

• Open Office Calc (.ods) spreadsheets
• Microsoft Excel (.xls) spreadsheets
• Comma-separated value (.csv) files

5.2 physiological recordings，生理记录

记录了脑电图（EEG）和外周生理信号的数据。提供三种格式：BioSemi 的原始未处理数据（.bdf 格式），Matlab 和 Python（numpy）格式的预处理数据（详见数据集描述）。

• Original BioSemi (.bdf) format (5.8GB)
• Preprocessed data in Matlab format (2.9GB)
• Preprocessed data in Python format (2.7GB)

有一些人报告使用下载管理器下载这些大文件时遇到了问题。如果你也面临这个问题，还有一个多卷版本的文件，其中文件被分成了100MB的部分。

5.3 video recordings, video recordings

This is a video recording of the forehead face of participants 1-22 in the experiment, in h264 format.

• Face videos in avi format (15.3GB)

If you have trouble downloading this large file using a download manager, there is a multi-volume version available where the file is split into 100MB sections. This makes it easier to download and manage files.