Unimore logo AImageLab

MultiSFace Dataset

About the dataset

The dataset consists of upper-body recordings of 31 subjects, with 16 different sequences for every subject. It contains variations in terms of age and sex and has also good coverage of accessories such as eyeglasses, sunglasses, hats, to simulate occlusions. We collected the data from different positions. In particular, the first 8 sequences of each subject are recorded from a distance of 1m, while the subject is positioned at 2.5m in the following 8 sequences. At both distances, 4 sequences are recorded without any occlusion and 4 sequences are recorded asking every subject to wear an accessory.

Acquisition devices

The sequences are recorded from multiple synchronized cameras:

  • Flir Boson 640, Thermal images 640x512px
  • PureThermal 2 (PT2) with Teledyne Flir Lepton 3.5, Radiometric thermal images, 160x120px
  • Pico Zense DCAM 710, RGB, depth and IR images, 640x360px
  • Pmdtec Pico Flexx, depth and IR images, 224x171px
  • Logitech Webcam, RGB images, 1920x1080px


For every subject, the dataset provides a folder containing a sub-folder for every data source (i.e. acquisition device) and a sub-sub-folder for every data type (e.g. depth, ir).

All RGB data and Boson thermal images are saved as video, while depth, IR, and PT2 thermal data are stored in .dat format. Python scripts to read data from .dat files and to extract all frames and save them as png images are provided.


To obtain a copy of the dataset, please fill in the licence agreement and send it to Stefano Pini and Guido Borghi.


We believe in open research and we are happy if you find this dataset useful. If you use it, please cite our paper.

  AUTHOR = {Pini, Stefano and Borghi, Guido and Vezzani, Roberto and Maltoni, Davide and Cucchiara, Rita},
  TITLE = {A Systematic Comparison of Depth Map Representations for Face Recognition},
  JOURNAL = {Sensors},
  VOLUME = {21},
  YEAR = {2021},
  NUMBER = {3},
  ISSN = {1424-8220},
  DOI = {10.3390/s21030944}