Overview
This research introduces EquiME, a synthetic micro-expression dataset designed to address limitations in existing micro-expression recognition (MER) datasets. These limitations primarily include small scale, restricted demographic coverage, and a narrow range of emotion labels. EquiME was constructed using an Action Unit (AU)-guided image-to-video generation process. The dataset encompasses 75,000 videos, derived from 15,000 distinct source face images, exhibiting five target emotions. Alongside the video data, EquiME includes automatically inferred demographic metadata and video-quality measurements.
Research Context
Micro-expression recognition faces challenges due to characteristics of current datasets. These datasets often feature a small overall scale, limiting the volume of available training data. Furthermore, they frequently exhibit narrow demographic coverage, potentially impacting the generalizability of models trained on them. An additional constraint is the limited set of emotion labels assigned within these datasets, which can hinder the development of nuanced emotion recognition systems. The creation of EquiME directly responds to these identified data limitations, aiming to provide a more expansive and diverse resource for MER research.
Approach
The core of the research approach involved developing EquiME through AU-guided image-to-video generation. This process leverages Action Units, which are fundamental movements of facial muscles, to synthesize micro-expressions programmatically. The dataset generation pipeline used a structured AU-conditioning mechanism. From an initial pool of 15,000 source face images, 75,000 micro-expression videos were generated. Each generated video corresponds to one of five target emotions. The dataset also integrates automatically inferred demographic metadata for each video and includes video-quality measurements to characterize the synthetic content.
Findings
EquiME comprises 75,000 synthetic micro-expression videos originating from 15,000 unique source face images, categorized across five target emotions. The dataset provides automatically inferred demographic metadata and video-quality metrics. Evaluation of EquiME involved assessing its characteristics using frame-pair similarity, spatial variation analysis, and no-reference perceptual-quality metrics. To assess its Utility for micro-expression recognition, cross-dataset MER experiments were conducted using SAMM and CASME II, which are existing micro-expression datasets.
The empirical results indicate that models trained on EquiME achieve competitive cross-dataset performance when evaluated on the SAMM and CASME II datasets. A further observation from these experiments is the comparatively low variation observed across the four evaluated architectures when trained with EquiME data.
Why This Matters
The development of EquipME addresses persistent challenges within micro-expression recognition research, specifically the scarcity of large-scale, demographically diverse, and richly labeled datasets. By providing a synthetically generated dataset that is significantly larger and broader in scope than many existing resources, EquiME offers a new avenue for training robust micro-expression models. The demonstrated competitive cross-dataset performance on established benchmarks like SAMM and CASME II suggests that synthetic data can serve as an effective substitute or supplement to real-world datasets, potentially accelerating progress in the field.
Potential Applications
EquiME could serve as a valuable resource for training and evaluating micro-expression recognition models, particularly in scenarios where the acquisition of large volumes of real micro-expression data is impractical or resource-intensive. Its size and structured AU-conditioning pipeline could also support the development of algorithms that are more robust to variations in facial characteristics and emotional expressions, owing to the dataset's inherent diversity in source images and controlled emotional synthesis. The dataset's inclusion of demographic metadata and video-quality metrics allows for further analysis of model performance across different groups and under varying synthetic data fidelities.