Facial Micro-Expression Recognition via 3D Residual Magnification

Abstract

Micro-expression recognition remains challenging because facial motions are subtle, brief, and vulnerable to appearance, illumination, and registration variations. This paper proposes a facial micro-expression recognition method based on 3D residual magnification. Detailed facial meshes are reconstructed frame by frame, and vertex-wise residuals relative to an approximately neutral first frame are magnified by a fixed factor while preserving mesh connectivity. The resulting meshes are rendered into detail-shape and depth sequences for spatiotemporal classification. LOSO experiments on SMIC-HS, CASME II, and SAMM show that FixedAmp achieves ACC/UAR/UF1 scores of 0.6929/0.5272/0.5221, 0.8157/0.4395/0.4200, and 0.8489/0.4183/0.4000, respectively. The motion-scaling ratios remain close to 2, with mean relative edge-length changes below 0.004, supporting FixedAmp as an interpretable baseline for subtle facial-motion enhancement.