Food Intake Cycle Dataset – v0

The Food Intake Cycle (FIC) dataset was created by the Multimedia Understanding Group towards the investigation of in-meal eating behavior. The dataset includes the accelerometer and gyroscope signals as captured by an off the shelf commercial wristband. Moreover, the FIC dataset also includes the hand-labeled start and end points for each every in-meal hand movement. The dataset is available for download by using the download link near the end of this page.

Description

For the purposes of this study, a total of 10 subjects were recorded while eating their launch at the university’s cafeteria. The total duration of the recording sums up to 109 minutes, with a mean duration of 13.2 minutes. The wristband used for the study was an off-the-shelf Microsoft Band 2, containing a triaxial accelerometer providing measurements in g units and gyroscope providing measurements in degrees per second (degrees/sec).

Furthermore, a Go Pro Hero 5 action camera was also used for generating ground truth annotations. It was mounted on a short tripod (approximately 23 cm height) placed on top of the table, while facing the subject’s upper torso and food tray.

All participants were free to select the food type of their choice. A typical recorded meal consists of a starter, a salad and a main course. Each participant was asked to wear the wristband encapsulating the sensors to the hand responsible for handling the fork. In order to synchronize the accelerometer and gyroscope streams with the camera feed, prior to starting his meal the subject was asked to clap his hands. This procedure enabled synchronization of the two streams with different timestamps, since the quick hand motion required for clapping exhibits an easily identifiable sharp peak in the acceleration signal. No other instructions were given to the subjects on how to eat their meals. As a result, the participants were free to engage in conversations with other individuals seated near them, use their cellphones and perform any other non meal-related activities. After the subject finished their meal, they were asked to perform a second clap marking the end of the collection procedure as well as verifying the initial synchronization offset.

Annotation

For all video recordings, the start and end points of all six micro-movements of interest were manually labeled. The micro-movements of interest include:

  • Pick food, hand manipulates a fork to pick food from the plate
  • Upwards, hand moves upwards, towards the mouth area
  • Downwards, hand moves downwards, away from the mouth area
  • Mouth, hand inserts food in mouth
  • No movement, hand exhibits no movement
  • Other movement, every other hand movement

The annotation process was performed in such a way that the start and end times of each micro-movement span the whole meal session, without overlapping each other.

Data Statistics

 
Label Intances Total duration (sec) Mean (± std) duration (sec)
P 727 1613.43 2.21 (± 1.71)
U 700 678.28 0.96 (± 0.58)
D 694 518.37 0.74 (± 0.57)
M 695 311.96 0.44 (± 0.17)
N 161 965.67 5.99 (± 5.71)
O 742 3837.92 5.17 (± 7.42)
 

In addition to the micro-movement statistics above, the average food intake cycle duration and the average distance between two consecutive food intakes were 5.39 (± 3.86) and 11.22 (± 8.79) seconds, respectively.

Download link

If you plan to use this dataset or other resources you’ll find in this page, please cite our paper:  
@inproceedings{kyritsis2017automated,
title={Automated Analysis of in Meal Eating Behavior using a Commercial Wristband IMU Sensor},
author={Kyritsis, Konstantinos and Tatli, Christina Lefkothea and Diou, Christos and Delopoulos, Anastasios},
booktitle={2017 39th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC)},
year={2017}
}
  The dataset can be downloaded by using this link. The downloaded zip file includes two folders, named Data and GT as well as Details text file.

Technical details

IMU Data

Each .mat file in /Data/ includes 2 N*4 matrices, namely acceleration and gyroscope. Variable N represents the number of samples in each session. The first column for both the acceleration and gyroscope matrices represents the timestamp in milliseconds. The next three columns represent the acceleration or rotation in the x,y and z axis respectively. The accelerometer signals are measured in g units whereas the gyroscope signals are measured in degrees/second. Both the acceleration and gyroscope signals are re-sampled to a sampling frequency of 100 Hz.

Groundtruth

Each .mat file in /GT/ includes a M*4 matrix, named GT. Variable M represents the number of annotated in-meal micro-movements for the specific subject. The second and third columns of each GT matrix represent the start and end points of the micro-movement respectively in milliseconds. The fourth column represents the observed micro-movement. The label to micro-movement mapping is provided below.

1 -> no movement, 2 -> upwards, 3 -> downwards, 4 -> pick food, 5 -> mouth, 6 -> other movement