KeyLemon FARGO Dataset
The KeyLemon FARGO Dataset is an RGB-D dataset containing recordings of 75 identities which were captured at three different sites during Summer 2016.
KeyLemon FARGO is no longer available for download
Each identity was captured twice with two sensor setups, mobile and laptop, and in three different lighting conditions, namely controlled, dark, and outdoor. In total, there are thus 12 recordings per identity.
The sensor is an Intel RealSense SR300.
A recording consists in approximatively 10 seconds of frontal face data, followed by 4 head movements (left, right, up, down). Although subjects were specifically asked to always keep an eye on the sensor, some have not strictly followed the instructions.
From the original raw and unsynchronized data, frames from the three camera channels have been synchronized at 10 frames per second.
Amongst the 75 identities, 20 are female and 55 male. The 33 subjects wearing glasses were asked to remove them for at least one recording. Most of the subjects were between 20 and 30 years old at the time of the recordings.
13 frames per recording were manually labeled using a custom landmarking scheme of 16 points:
- eyebrows: left and right ends,
- eyes: left and right corners,
- nose: nose tip,
- mouth: left and right corners, top and bottom of the lips,
- chin: chin tip, left and right chin border at ear's height.
The 13 frames were selected as follows:
1. first frame of the sequence,
2. extreme left head position,
3. extreme right head position,
4. extreme up head position,
5. extreme down head position,
6-7. two frames between 1 and 2,
8-9. two frames between 2 and 3,
10-11. two frames between 3 and 4,
12-13. two frames between 4 and 5.
A recording consists in approximatively 10 seconds of frontal face data, followed by 4 head movements (left, right, up, down). Although subjects were specifically asked to always keep an eye on the sensor, some have not strictly followed the instructions.
From the original raw and unsynchronized data, frames from the three camera channels have been synchronized at 10 frames per second.
Amongst the 75 identities, 20 are female and 55 male. The 33 subjects wearing glasses were asked to remove them for at least one recording. Most of the subjects were between 20 and 30 years old at the time of the recordings.
13 frames per recording were manually labeled using a custom landmarking scheme of 16 points:
- eyebrows: left and right ends,
- eyes: left and right corners,
- nose: nose tip,
- mouth: left and right corners, top and bottom of the lips,
- chin: chin tip, left and right chin border at ear's height.
The 13 frames were selected as follows:
1. first frame of the sequence,
2. extreme left head position,
3. extreme right head position,
4. extreme up head position,
5. extreme down head position,
6-7. two frames between 1 and 2,
8-9. two frames between 2 and 3,
10-11. two frames between 3 and 4,
12-13. two frames between 4 and 5.
CITATION: G. Heusch, T. de Freitas Pereira and S. Marcel, "A Comprehensive Experimental and Reproducible Study on Selfie Biometrics in Multistream and Heterogeneous Settings", IEEE Transactions on Biometrics, Behavior and Identity Science, 2019.