Indoor Action Dataset

Indoor Action Dataset

We are releasing the Indoor video dataset for action recognition presented in [1]. This dataset includes recordings of several common scenes for indoor daily life activities. It was built recording samples of an average of 10-minute videos at 25 fps with people carrying out indoor actions. Afterwards, the videos were manually curated and sampled into 3-to-5-second labeled clips. In total, five different subjects collaborated in the recordings, in a variety of indoor scenarios such as bedrooms, kitchens, or living rooms.

The dataset includes samples of ten different classes of activities such as cleaning, eating, sitting down, standing up, blowing nose, walking, watching tv as well as classes that represent potentially risky situations such as falling down or lying on the floor. Additionally, we included a no-action class with indoor spaces where no activity is carried out or no one is on the scene.

In the following table we provide information about how many clips were recorded by each subjects.

Subject ID
Gender
Train
Val.
Test
Total

S0
M
406
92
123
621

S1
M
32
0
85
117

S2
M
70
25
23
124

S3
F
70
0
63
133

S4
F
95
70
68
233

In addition, the Table below shows the number of samples retrieved per class or daily life activity.

Action name
Train
Val.
Test
Total

blowing nose or sneezing
38
10
15
63

cleaning
55
17
34
106

eating
57
18
32
107

falling down
34
11
16
61

lying on the floor
53
12
36
101

sitting down
66
22
29
117

standing up
95
30
41
166

walking
158
34
90
282

watching tv
60
19
43
122

no-action
57
14
32
103

How-to-use?
First, download the video Indoor Action Dataset running setup_video_data.sh (Available upon publication). This will download a zip file with the dataset and will unzip it into three folders: train, validation and test.

git clone https://github.com/DaniDeniz/IndoorActionDataset.git
cd IndoorActionDataset
chmod +x setup_video_data.sh && ./setup_video_data.sh

The full zip file is also uploaded to this repository.

Full lenght videos were manually cropped into individual clips, and then this clips were assigned to a set (train, validation or test). Bear in mind that all clips for each video are assigned to the same set, avoiding having train, validation and test samples with similar propierties. We are providing the clips with HD (720p) resolution.

The name of each clip is as follows:

20210114 _ 150036 − 151348 [ 150942.072000 , 150950.572000 ] ( L i v i n g R o o m 1 − S 0 )

Which represents:

d a t e ( Y Y Y Y M M D D ) _ s t a r T i m e V i d e o ( H H M M S S ) − f i n i s h T i m e V i d e o ( H H M M S S )

[ s t a r t T i m e C l i p ( H H M M S S . f ) , f i n i s h T i m e C l i p ( H H M M S S . f ) ] ( S c e n a r i o I D − S u b j e c t I D )

Note that with Video we refer to the original video of continuous recording, and with Clip we refer to each of the samples extracted from each video. Therefore, every clip recorded in the same scenario by the same subject at the same Video time means that these clips were retrieved from the same video. The start and finish Clip time points out when the action occurred in the recorded Video.

Citation
Please cite this paper in publications carrying out work using this video database:

[1] D. Deniz, E. Ros, E. M. Ortigosa, F. Barranco. "Optimized edge-cloud system for activity monitoring using knowledge distillation" in Electronics, 13 (23), 2024.

@article{deniz2024optimized,
title={Optimized Edge-Cloud System for Activity Monitoring Using Knowledge Distillation},
author={Deniz, Daniel and Ros, Eduardo and Ortigosa, Eva M and Barranco, Francisco},
journal={Electronics},
volume={13},
number={23},
pages={4786},
year={2024},
publisher={MDPI}
}

License
BSD 3-Clause license

Acknowledgements
This work was supported by the National Grant PID2022-141466OB-I00 funded by MICIU/AEI/10.13039/501100011033 and by ERDF/EU.