Research

  • Research at CEA LIST : Examples of Human Behaviour research topics

    Daily activity understanding

    Affective Computing
    User / Environment Interactions

  • Realistic datasets

DAHLIA (DAily Human Life Activity)

DAHLIA dataset is devoted to human activity recognition, which is a major issue for adapting smart-home services such as user assistance.

DAHLIA has been realized in Mobile Mii Platform by CEA LIST, and has been partly supported by ITEA 3 Emospaces Project (https://itea3.org/project/emospaces.html)

DALHIA has been published at the 12th IEEE Conference on Automatic Face & Gesture Recognition, Washington DC, USA, May 31th – 3rd June, 2017. See Paper here

Videos were recorded in realistic conditions, with 3 Kinect v2 sensors located as they would be in a real context. The long-range activities were performed in an unconstrained way (participants received only few instructions), and in a continuous (untrimmed) sequence, resulting in long videos (40 min in average per subject). Contrary to previously published databases, in which labeled actions are very short and have low-semantic level, this new database focuses on high-level semantic activities such as « Preparing lunch » or « House Working ».

DAHLIA Datafile Package:

The dataset is devided into two sets:
– set1: 25 examples
– set2: 26 examples

The tree structure and content of directories is done as below:

set1/
S01_A1/
Label/
S01_A1_K1/
S01_A1_K2/
S01_A1_K3/
Color/
Depth/
Time/
Body/
BodyIndex.7z

In a set, examples are ordered by subject and acquisition number.
In each examples, you will find the directories for each stream of the three cameras and the Label directory which is common to all views.

  • Label : The xml files with activity label for each frames.
    Activities are : Cooking (id: 1) – Laying table (2) – Eating (3) – Clearing Table (4) – Washing dishes (5) – Housework (6) – Working (7) – Neutral (0)
  • Color : The RGB avi video file
  • Depth : The jpg2000 files corresponding to depth map
  • Time : The timestamp files
  • Body : The csv files with skeleton 3D coordinates.
    Fileskeleton.csv: expressed in depth map referential. FileskeletonSkSpace.csv: expressend in real world referential.
  • BodyIndex.7z : (when extracted,) The body index binaries files which can be decoded in tif files thanks to the C++ software: https://github.com/CEA-LIST/DAHLIA-BodyIndex
    You will then obtain a black and white image correponding to the body index returned by the Kinect v2.

Benchmark on DAHLIA:

Benchmark               
Skeleton based
View 1View 2View 3Multi-ViewCross-View
FA1F-ScoreIoUFA1F-ScoreIoUFA1F-ScoreIoUFA1F-ScoreIoUFA1F-ScoreIoU
DOHT0.600.580.420.630.600.440.730.710.560.650.640.480.340.310.19
ELS0.180.180.110.270.260.160.520.550.390.310.320.21
Color based
View 1View 2View 3Multi-ViewCross-View
FA1F-ScoreIoUFA1F-ScoreIoUFA1F-ScoreIoUFA1F-ScoreIoUFA1F-ScoreIoU
DOHT0.800.770.640.810.790.660.800.770.650.850.820.71
Max Subgraph Search0.250.150.180.100.440.31
Sub Activity0.850.810.730.870.820.750.820.760.690.850.820.71

If you have published results of activity recognition on the DAHLIA dataset, please send us an eMail (dahlia.mobilemii@cea.fr) with a link to your official publication in order to update the benchmark table.

Dataset Request Formular:


Please download the formular, fulfill, sign and send it by eMail to : dahlia.mobilemii@cea.fr
Once we received your fulfiled request, a login/password will be sent to you back and you will be able to connect in section Member Login. Therefore, links below will be available.