AffectNet+ is a large-scale extension of the original AffectNet dataset that introduces soft emotion labels and rich metadata for approximately one million facial images. Using an ensemble of state-of-the-art expression recognition models, AffectNet+ generates plausible emotion probability vectors for all samples, explicitly modeling annotation uncertainty. The dataset also provides automatically extracted demographic attributes, facial embeddings, and landmark structures, enabling research on fairness, robustness, and dataset bias. Models trained on AffectNet+ demonstrate improved accuracy, calibration, and reduced demographic bias compared to those trained on hard-label annotations, establishing AffectNet+ as a new benchmark for facial-expression recognition.

AffectNet and AffectNet+ are large-scale facial expression datasets designed to advance affective computing research across both categorical emotion recognition and continuous valence–arousal modeling. Together, they comprise over one million facial images collected from the internet using 1,250 emotion-related keywords in six languages, with approximately 440,000 images manually annotated in the original AffectNet for seven discrete facial expressions and valence–arousal intensity.

AffectNet+ builds on this foundation by reprocessing the entire dataset to provide improved annotation reliability through soft emotion labels and enriched metadata, while preserving the original scale and diversity.

DISFA and DISFA+ are benchmark databases used to detect the intensity of facial Action Units (AUs) from spontaneous facial expressions. DISFA+ includes additional annotations for valence and arousal, which are useful for modeling continuous dimensional affect.

These databases have supported numerous research studies on automatic Action Unit detection and the interpretation of facial expressions in the context of emotional responses.

RyanSpeech is a high-quality male speech corpus designed for text-to-speech (TTS) research. Unlike many publicly available TTS corpora, it features over 10 hours of a professional male voice actor's speech recorded at 44.1 kHz, derived from real-world conversational settings. Optimized for developing TTS systems in real-world applications, RyanSpeech has been used to train state-of-the-art speech models, achieving a 3.36 mean opinion score (MOS) in its best model.

Fingerprint-based positioning is emerging as a reliable alternative to GPS and cellular localization in urban areas. However, the lack of public datasets limits research in this field. To address this, we introduce OutFin—a comprehensive outdoor fingerprint dataset collected using two smartphones. It includes WiFi, Bluetooth, and cellular signals, along with sensor data from the magnetometer, accelerometer, gyroscope, barometer, and ambient light sensor. Data was gathered from 122 reference points across four distinct sites, each with unique GNSS visibility and layout. Prior to release, OutFin underwent rigorous testing to ensure data quality, making it a valuable resource for advancing fingerprint-based positioning research.