BlanketSet - a clinical real word action recognition and qualitative semi-synchronized MoCap dataset

BlanketSet consists of 405 RGB-IR-D videos recorded with an Azure Kinect of people in a hospital bed performing 8 different movement sequences with different blankets, speeds, lighting conditions, hand positions, and blanket initial positions. Each recording was repeated 3 times with the 3 different blanket initial positions (body fully covered, only bottom half covered, and no blanket). The dataset was acquired with an Azure Kinect at the Epilepsy Monitoring Unit (EMU) of the University Hospital Center of São João in several sessions over the course of several months. Covid restrictions changed while the dataset was acquired, this led to the participants wearing masks in some recordings, but not in others. One person participated both before and after the restrictions changed, and therefore both with and without a mask.

To each participant, the data to be recorded was explained, as well as their privacy rights, before they each signed an informed consent form in which they agreed to the recording and publication of their data in the dataset.

To request access to this dataset please fill out the form below (BlanketSet request form) and send it to: joao.a.castro@inesctec.pt

If you use BlanketSet you agree to cite the corresponding paper: João Carmona, Tamás Karácsony, and João Paulo Silva Cunha. "BlanketSet--A clinical real word action recognition and qualitative semi-synchronised MoCap dataset." arXiv preprint arXiv:2210.03600 (2022).

Data ja resurssit

Lisätietoja

Kenttä Arvo
Laatija João Carmona
Viimeksi päivitetty toukokuuta 3, 2024, 15:03 (UTC)
Luotu lokakuuta 21, 2022, 14:15 (UTC)
Blanket Originally 3 different blankets of different thicknesses, colors, and weights (black heavy, white light, and green heavy) were alternated between. In the recordings after the restrictions were lifted, different blankets were used (grey heavy, white light, orange heavy).
Blanket Initial Position There were 3 different initial positions for the blanket: no blanket, fully covering the subject, and pulled down so that only about the feet and lower legs were covered.
Citation Carmona, J., Karacsony, T., & Cunha, J. P. (2022). BlanketSet - a clinical real word action recognition and qualitative semi-synchronized MoCap dataset [Data set]. INESC TEC. https://doi.org/10.25747/QMYY-GV63
Contributor Tamás Karacsony, João Paulo Cunha
DOI https://doi.org/10.25747/qmyy-gv63
Muoto .mkv, .csv
Hand Position Four different hand positions were included: relaxed (1), fingers spread (2), fingers stretched touching (3), and closed fist (4).
Instrument RGB-IR-D camera
Instrument Name Microsoft Azure Kinect
Kieli EN
Lighting 4 different lighting setups were used: natural lightning controlled by having the blinds up or down in the window next to the bed, and having the lights on or off.
Movement Sequences Each movement sequence consists of switching between a few positions repeatedly, each recording contains 20 position changes. The emphasis was on ensuring that the motions were repeated properly over ensuring they were done perfectly as described, it was deemed more important to have consistency within each set of recordings than within the dataset as a whole. In order to cover all the variations of the controlled variables, each sequence was repeated 9 times, once for every combination of BIP and blanket. The semi-controlled variables were changed with the blanket, such that for every recording with one BIP, there be recordings with the other BIPs and no other changed variables.
Parameters In order to control for as many variables as possible within reason, the different variables of the dataset were split into three categories: controlled variables, semi-controlled variables, and uncontrolled variables. Controlled variables: were controlled to cover the dimension space as evenly as possible. Semi-controlled variables: were randomly sampled from a uniform distribution so as to avoid correlations between them and other variables. Uncontrolled variables: were deemed not a priority to control in this dataset, although they could affect the performance of systems trained on the dataset. These variables include the participants´ gender, ethnicity, body shape, clothing, hair color and length, and the time of day.
Temporal Coverage April-July 2022