Over the past few decades, human recognition has been extensively studied. Most of the research concentrates on typical physiological traits, e.g., face, fingerprint, voice, etc., which are usually distinctive for different individuals but difficult to semantically describe. The recent advances have aroused the emerging research in soft biometric attributes, e.g., gender, age, accent, etc., which are not necessarily unique but have semantic interpretations. Such soft biometric attribute offers middle-level characteristics of a person to bridge the gap between low-level machine features and high-level human descriptions. The produced middle-level characteristics are particularly useful in large-scale biometric identification applications, such as human-machine interaction, visual tagging/indexing, and person re-identification. Soft biometric learning and recognition from big data thus has become a very active inter-disciplinary research area, involving computer vision, machine learning and biometrics. The goal of the workshop is to disseminate recent research findings for researchers on a focused platform, foster in-depth discussion of technical solutions, identify application opportunities for large-scale soft biometrics, and explore potential collaborations.