Bio-perception oriented intelligent computing (BPOIC) is an interdisciplinary field spanning computer science, psychology, cognitive science, and medical science. BPOIC is essential for deep understanding of human by machines covering visible attributes (e.g., demographic expression) to latent characteristics (e.g., psychological and physiological). BPOIC is challenging because of data source diversity, modality heterogeneity, data or label missing, data imbalance, etc. We aim at the research and development of methods for modal image/video enhancement, visual attribute learning, intelligent healthcare, and representation learning, etc. The goal is to establish the capability of robust bio-perception under unconstrained scenarios.
1. BPOIC problem: RGB-D face recognition, 3D face modeling, biometrics and soft-biometrics, visual attribute learning, presentation attack detection, emotion recognition; rPPG based physiological measurement, medial image analysis (XRay, CT, MRI, Echocardiography, etc.).
2. BPOIC methodology: Multimodal data enhancement and processing based on self-supervised learning, cross-modality correlation, and domain knowledge; multi-modality complementary feature representation learning and cross-modality correlation modeling; spatial-temporal representation learning; multi-task leaning via task correlation and heterogeneity modeling; learning with missing and noisy labels.