The AISL HDPIPL Database is a collection of a large-scale human depth images with body part labels.
Currently, we have the datasets for the following three cases:

1. Sitting poses viewed from the front
2. Recumbent poses viewed from various directions
3. Sitting poses with a table and a chair viewed from various directions

In the aging society, the demands for service robots which take care of the elderly in care houses and at home. Human pose recognition is one of the necessary functions of such robots, which contributes to an accurate care and diagnosis or anomaly detection. Many machine learning-based methods have been proposed for human pose estimation in color image. A drawback of color-based estimation is that it is sensitive to changes of illumination, clothing and skin color. They could also raise a privacy issue. Use of depth images is a promising alternative to solve these problems. However, this leads to another big problem, that is, to construct a large dataset of annotated depth images. Since the annotating real depth images is difficult, we developed an efficient generation method of large-scale dataset of human depth images with body part labels by using computer graphics models.
In this page, we provide dataset of human depth images with body part labels by using our method.

Each dataset contains two type data: depth, label.
We normalized the size to 212 x 212 pixels.
Depth
These data are xml file which contaion human depth information.
We normalized depth value from range [0mm, 2000mm] to [0, 1].
So, if you want to real depth range data, please multipule it by value of 2000.
Label
These data are png image which contain human body label information.
We defined eleven parts and one background:

The AISL HDIBPL (Human Depth Images with Bosy Part Labels) Database is now made available for research purpose only.
The researcher(s) is free to use the AISL HDIBPL database by obeying and agreeing the following restrictions on the AISL HDIBPL database:

1. The database will not be further distributed, published, copied, or further disseminated in any way whether for profit or not.
2. All the images will be used for the purpose of scientific researches only. The AISL HDIBPL database, in whole or in part, will not be used for any commercial purpose in any form.
3. All technical papers, documents and reports which use the AISL HDPIBPL database will acknowledge the use of the database by a citation to
“K. Nishi and J. Miura, Generation of Human Depth Images with Body Part Labels for Complex Human Pose Recognition, Pattern Recognition (under review)”.

1. K. Nishi and J. Miura, A Head Position Estimation Method for a Variety of Recumbent Positions for a Care Robot, Proc. 2015 Int. Conf. on Advanced Mechatronics (ICAM-2015), Tokyo, Japan, Dec. 2015.

Please fill in all fields and check the box below to continue download the database.