GAMEEMO: Facial & EEG Emotion Datasets

• GAMEEMO is a dual-sense emotion dataset comprising both a game-based facial image collection (GaMo) and an EEG dataset with continuous self-reported ratings.
• The facial subset uses an interactive game framework to capture in-the-wild expressions with automatic labeling and balanced class distributions.
• The EEG subset records 14-channel brain signals during serious games, enabling subject-independent classification across various affective states.

The label GAMEEMO is used in two distinct senses in the affective computing literature represented here. In the facial-expression literature, GAMEEMO is an alias for the GaMo (game-based emotion) dataset, introduced in “Towards an ‘In-the-Wild’ Emotion Dataset Using a Game-based Framework” and collected from a web game that automatically labels webcam facial images (Li et al., 2016). In later EEG-based work, GAMEEMO denotes a dataset of 14-channel EEG recordings with continuous self-reported emotion ratings from serious games, used for subject-independent emotion classification across binary valence, multi-class, and fine-grained multi-label settings (Rehman et al., 28 Aug 2025). The shared name therefore designates two different resources, one image-based and one EEG-based, with different annotation schemes, collection protocols, and intended uses.

1. Naming, provenance, and scope

The facial-expression resource is officially named GaMo (game-based emotion) dataset. In some citations and follow-on work it is also referred to as GAMEEMO; in the 2016 paper, however, the authors consistently use GaMo, and “GAMEEMO” is explicitly described as an alias for the same dataset. That paper is authored by Wei Li, Farnaz Abtahi, Christina Tsangouri, and Zhigang Zhu, with affiliations at the Department of Electrical Engineering, CUNY City College, the Department of Computer Science, CUNY Graduate Center, and the Department of Computer Science, CUNY City College, New York, USA. Data collection was conducted under IRB approval, and all users signed consent (Li et al., 2016).

The EEG resource described in the later literature is a different dataset. In “Spatiotemporal EEG-Based Emotion Recognition Using SAM Ratings from Serious Games with Hybrid Deep Learning,” GAMEEMO consists of 14-channel EEG recordings and continuous self-reported emotion ratings from 28 subjects across four emotion-inducing gameplay scenarios. That paper references GAMEEMO (Alakus et al., 2020, Biomedical Signal Processing and Control, 60:101951), but does not provide a direct repository link or full acquisition metadata (Rehman et al., 28 Aug 2025).

Usage of the name	Modality	Core annotation scheme
GaMo / GAMEEMO alias	Webcam facial images	Target emotion shown in the game at capture time
GAMEEMO in EEG work	14-channel EEG	Continuous ratings for boring, horrible, calm, funny

This dual usage matters because the two datasets are not interchangeable. A plausible implication is that references to “GAMEEMO” should be interpreted only in conjunction with modality, citation, and task definition.

2. GaMo as a game-based facial emotion dataset

GaMo was proposed to create an “in-the-wild” facial emotion dataset with a large number of balanced samples. The motivating problem was that existing facial-emotion datasets were either constrained—for example, in laboratory settings with controlled lighting and pose, such as CK+, MMI, and DISFA—or else collected from media or the web, such as CIFE and EmotiW, but heavily imbalanced, especially with over-representation of happiness and sadness and under-representation of disgust and fear. The proposed remedy was a game-based framework for low-cost crowdsourcing with built-in supervision, class-balance control, and automatic labeling (Li et al., 2016).

The framework has three components: a game engine, a game interface, and a data collection and evaluation module. The game engine is a CNN emotion classifier trained on CIFE and deployed server-side because of computational demand. For each input image it returns a 7-way probability vector. The game interface is a web-based “Tower Defense”-style game in which emotion icons, described as “bombs,” drop from the top of the screen. The player’s webcam video is displayed; to defuse a bomb and score, the player imitates the displayed emotion, and the engine must recognize the match before the bomb reaches the ground. Each session has five lives (“hearts”), and emotion targets are sampled with equal probability across the seven emotions to promote class balance (Li et al., 2016).

Two gameplay modes are described. In the general version (recognition), the browser receives the server’s probability vector $p$ and saves an image if the top prediction matches the current target. In the customized version (verification), the problem is simplified to checking whether the current image expresses the known target emotion, using per-emotion thresholds because some emotions are harder to mimic. The decision rule is

$$\text{accept emotion } e \text{ if } p_e \ge \tau_e,$$

where $\tau_e$ is a predefined threshold for class $e$. The customized version also allows each registered user to build seven emotion templates. The server extracts CNN features for each template and classifies incoming frames by nearest-template comparison using the $L_2$ distance,

$d(x, t_k) = \|x - t_k\|_2,$

choosing the nearest template as the recognized class. The paper states that this personalization improves user scores and encourages longer play, thereby increasing data volume (Li et al., 2016).

Operationally, images are transmitted at 1 Hz, both to limit server load and to allow time for expression formation. Face detection is performed during template registration; if no face is detected, the user is prompted to recapture. Server inference time is reported as about 200 ms per image, which the authors state keeps gameplay smooth (Li et al., 2016).

3. Dataset composition, annotation, and empirical behavior of GaMo

GaMo contains 15,455 images, collected in the first month of operation. The dataset covers seven basic emotion classes: Angry, Disgust, Fear, Happy, Neutral, Sad, and Surprise. The per-class counts are 1,945 Angry, 1,838 Disgust, 1,586 Fear, 3,185 Happy, 2,741 Neutral, 1,898 Sad, and 2,262 Surprise. More than 100 users played the general version, 74 users tried the customized version during the first month, and a separate five-user study—3 male, 2 female—was used for engine evaluation. The exact number of unique contributors to the dataset is not specified, but collection involved college students across both versions (Li et al., 2016).

The images are described as “in-the-wild” because they were collected with built-in or external webcams, under natural lighting, with varied viewing angles and distances, and in users’ own environments. The resulting images include both explicit and subtle expressions, as well as occasional occlusion and non-frontal poses. Labels are assigned automatically: every saved image inherits the target emotion shown in the game at capture time. In the general version, saving requires that the engine’s predicted class match the target; in the customized version, saving requires threshold verification or nearest-template agreement. The authors report that no manual cleanup was performed and that random spot checks did not reveal labels “very off” from the true expressions (Li et al., 2016).

The CNN used as the game engine is based on the “Alex” model (AlexNet), pretrained on ImageNet and fine-tuned on CIFE for seven emotions. The architecture is modified so that fc7 has 2048 kernels and fc8 has 7 outputs, with earlier layers also updated during fine-tuning. After GaMo was collected, a second CNN was trained by further fine-tuning the same backbone on GaMo (Li et al., 2016).

Evaluation is reported both as self-test and cross-dataset test. The CIFE-trained model achieved average accuracy 0.74 on CIFE, with per-class accuracies Angry 0.68, Disgust 0.29, Fear 0.44, Happy 0.87, Neutral 0.75, Sad 0.79, and Surprise 0.73. The GaMo-trained model achieved average accuracy 0.64 on GaMo, with Angry 0.65, Disgust 0.57, Fear 0.52, Happy 0.71, Neutral 0.71, Sad 0.64, and Surprise 0.65. The paper emphasizes that the GaMo result is more balanced across classes. In cross-dataset evaluation, the CIFE-trained model tested on GaMo dropped to average 0.21, with especially low performance on Angry 0.03, Disgust 0.02, and Happy 0.03. By contrast, the GaMo-trained model tested on CIFE reached average 0.50, with Happy 0.80 and more moderate transfer across the other classes. The authors interpret this asymmetry as evidence that GaMo images capture more subtle, realistic expressions, while CIFE often contains exaggerated expressions, and that training on GaMo yields a more robust detector. A five-user gameplay study further showed higher average scores with the GaMo engine than with the CIFE-trained engine (Li et al., 2016).

The paper does not report official train/validation/test splits. Recommended use is therefore framed in terms of fine-tuning CNNs on GaMo for robust recognition of subtle expressions and using cross-evaluation against CIFE or similar datasets to assess generalization (Li et al., 2016).

4. GAMEEMO as an EEG dataset derived from serious games

In the EEG literature represented here, GAMEEMO is a dataset of 28 subjects, each recorded with 14 EEG channels during four emotion-inducing gameplay sessions. Each session is paired with continuous self-reported ratings for four affective states: boring, horrible, calm, and funny. The paper states both “scale 1–10” and “continuous scale 0–10”, and it does not explain the discrepancy. The exact channel names, montage, hardware/device, sampling rate, reference scheme, number of trials per session, within-session structure, duration per scenario, total recording time, and environment setup are not provided (Rehman et al., 28 Aug 2025).

The title references Self-Assessment Manikin (SAM), but the paper does not provide details on the classical SAM pictorial items or on any mapping from valence/arousal/dominance to the four emotion ratings. Instead, the work uses direct per-emotion intensity ratings for boring, horrible, calm, and funny. The paper also mentions “synchronized PDF-based self-reports,” but the exact timing of rating—whether during or immediately after each scenario—the precise synchronization mechanism, and the rating frequency are not specified (Rehman et al., 28 Aug 2025).

Three task formulations are defined from these ratings. For binary valence classification, funny and calm are treated as positive, while boring and horrible are treated as negative. With per-window or per-segment ratings $s_{\text{funny}}$, $s_{\text{calm}}$, $s_{\text{boring}}$, and $s_{\text{horrible}} \in [0,10]$, the paper defines

$$\text{accept emotion } e \text{ if } p_e \ge \tau_e,$$0

and

$$\text{accept emotion } e \text{ if } p_e \ge \tau_e,$$1

For multi-class emotion classification, the label is the most dominant emotion,

$$\text{accept emotion } e \text{ if } p_e \ge \tau_e,$$2

The paper does not provide a tie-breaking strategy. For fine-grained multi-label representation, each emotion rating is binned into ordinal classes. However, the paper states both “10 ordinal classes” and “11 ordinal bins (0–10)”, without specifying exact bin edges or encoding details (Rehman et al., 28 Aug 2025).

5. Preprocessing, feature extraction, and modeling for EEG GAMEEMO

The EEG pipeline uses temporal window segmentation with a sliding window length of $$\text{accept emotion } e \text{ if } p_e \ge \tau_e,$$3 and 50% overlap. All 14 channels are used. The paper mentions padding to maintain uniform segment lengths, but does not specify whether this is zero-padding, reflection, or another method. It also does not specify band-pass ranges, notch filters, artifact removal procedures such as ICA or EOG regression, or any re-referencing step (Rehman et al., 28 Aug 2025).

Normalization is performed by z-score normalization,

$$\text{accept emotion } e \text{ if } p_e \ge \tau_e,$$4

where $$\text{accept emotion } e \text{ if } p_e \ge \tau_e,$$5 is the feature value, $$\text{accept emotion } e \text{ if } p_e \ge \tau_e,$$6 is the mean, and $$\text{accept emotion } e \text{ if } p_e \ge \tau_e,$$7 is the standard deviation computed over the training set. The paper presents this generally and does not specify whether the statistics are per-subject or global (Rehman et al., 28 Aug 2025).

Feature extraction combines time-domain and frequency-domain descriptors. The paper lists mean, variance, standard deviation, and entropy as time-domain features, and also mentions skewness and kurtosis in the dataset section. Frequency-domain features are based on FFT-based power spectral density and band powers across $$\text{accept emotion } e \text{ if } p_e \ge \tau_e,$$8, $$\text{accept emotion } e \text{ if } p_e \ge \tau_e,$$9, $\tau_e$0, and $\tau_e$1, using the generic band-power definition

$\tau_e$2

where $\tau_e$3 is the spectral density and $\tau_e$4 defines the band. The exact band boundaries are not provided. Gamma is mentioned in related work on other datasets, but not used explicitly in the paper’s own feature list (Rehman et al., 28 Aug 2025).

The evaluated models include the classical baselines Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), and XGBoost (XGB), alongside the deep architectures LSTM, LSTM-GRU, and CNN-LSTM; the algorithm section also writes the latter as LSTM-CNN. Architectural details such as the number of layers, hidden units, and kernel sizes are not provided. The stated optimization setup uses AdamW, initial learning rate $\tau_e$5, cosine annealing with warm-up, dropout 0.3, weight decay $\tau_e$6, gradient clipping norm $\tau_e$7, 100 epochs, and batch size 32. The evaluation uses a subject-independent 80:20 split, with subjects in validation or test unseen during training (Rehman et al., 28 Aug 2025).

The paper states cross-entropy (CE) for binary and categorical settings and also states “Multi-class: BCE” in Algorithm 1, noting an unconventional pairing that is not further clarified. This inconsistency is part of the reported methodological record rather than a resolved design choice (Rehman et al., 28 Aug 2025).

6. Reported results, limitations, and interpretive issues

For the EEG dataset, the reported best model is LSTM-GRU. In binary valence classification, it achieves F1-score = 0.932 and accuracy = 93.3%. In multi-class classification, its per-class metrics are boring: F1 0.939, accuracy 0.955, precision 0.942; horrible: F1 0.943, accuracy 0.953, precision 0.948; calm: F1 0.927, accuracy 0.938, precision 0.929; and funny: F1 0.935, accuracy 0.934, precision 0.934, with macro average F1 0.936, accuracy 0.945, and precision 0.938. In the fine-grained multi-label setting, LSTM-GRU reaches accuracy = 90.6%, F1-score = 0.909, and precision = 0.909. The corresponding Random Forest results are lower: approximately 85% accuracy in binary valence, macro average F1 0.758 and accuracy 0.796 in multi-class, and accuracy 0.79, F1-score 0.79, precision 0.79 in the fine-grained multi-label setting. The paper describes the LSTM-GRU confusion matrices as having a sharper diagonal and fewer off-diagonal errors, while Random Forest is said to struggle more with overlapping classes and with confusion between “boring” and “horrible” in the fine-grained setting (Rehman et al., 28 Aug 2025).

The paper attributes the performance advantage of LSTM-GRU to the ability of a hybrid temporal architecture to capture spatiotemporal dependencies in the 500 ms overlapping windows, and to the effect of feature-rich preprocessing and z-score normalization in reducing inter-subject variability. It does not provide ablation or sensitivity analyses, although it reports training curves with stable convergence and no obvious overfitting (Rehman et al., 28 Aug 2025).

Both dataset lines come with explicit limitations. For GaMo, the paper notes that class imbalance is reduced but not eliminated, that automatic labeling can introduce potential noise, that the total size of more than 15,000 images is still modest relative to modern deep learning needs, and that early data were collected primarily from college students willing to play an online game, with possible effects from self-selection and hardware variability. The use of customized templates may also encode user-specific expression styles (Li et al., 2016). For the EEG GAMEEMO dataset as described in the later work, limitations include the relatively small sample of 28 subjects, possible noise or bias in self-reported emotion intensities, the unresolved 0–10 versus 1–10 scale ambiguity, the inconsistency between 10 and 11 ordinal bins, and the absence of crucial acquisition details such as device, sampling rate, montage, reference, filtering, and artifact removal (Rehman et al., 28 Aug 2025).

Taken together, the two usages of GAMEEMO correspond to two different methodological traditions within affective computing. The facial-image resource emphasizes interactive game-based collection, balanced class design, and automatic labeling under in-the-wild visual conditions. The EEG resource, as used in later modeling work, emphasizes serious-game induction, continuous self-report, and multigranularity affective classification from 14-channel electrophysiological data. This suggests that the most important prerequisite for correct interpretation is not the shared name itself, but the associated modality, paper citation, and label definition.