WIMID: Wrist IMU Meals and Intakes Dataset

WIMID (Wrist IMU Meals and Intakes Dataset) is a free-living wearable sensor dataset intended for machine-learning research on eating behavior recognition. It consists of triaxial accelerometer and gyroscope recordings acquired from a smartwatch worn on the dominant hand during participants’ everyday activities, covering both eating and non-eating periods. The dataset was recorded from 11 participants and spans approximately 528 hours of data, including 159 meals. For 75 meals, participants additionally provided self-reported counts of dietary intakes. Information about how dietary intakes were performed (e.g., by hand, fork, spoon, or chopsticks) is also provided. Liquid consumption is excluded from the definition of eating. IMU data were recorded at a nominal sampling rate of 100 Hz. Due to minor variations in the actual recording frequency, signals were linearly interpolated to 100 Hz and subsequently smoothed using median filtering. The dataset provides the resulting preprocessed IMU recordings. The data were collected by researchers at the Jožef Stefan Institute, Slovenia.