InriaLHB – Inria Labeled Human Body

We present a new dataset of annotations for aligning a Point Distribution Model onto the human body. The corresponding research is described in the paper O. Huynh and B.Stanciulescu, “Clustering and Classifying Deformations for Shape Regression applied to the Human Body”, in ISPA 2015. The InriaLHB uses the pedestrian images from the Inria Person Dataset. Annotated shapes correspond to a standing person in front or back view.

Some annotated samples of the dataset

Model

The Point Distribution Model employed is defined with the following 29 landmarks :

0 – Top_Left_Head

1 – Top_Right_Head

2 – Bottom_Right_Head

3 – Bottom_Left_Head

4 – Left_Shoulder

5 – Left_External_Elbow

6 – Left_External_Wrist

7 – Left_Internal_Wrist

8 – Left_Internal_Elbow

9 – Left_Axilla

10 – Left_Pelvis

11 – Left_External_Knee

12 – Left_External_Ankle

13 – Left_Foot

14 – Left_Internal_Ankle

15 – Left_Internal_Knee

16 – Center_Pelvis

17 – Right_Internal_Knee

18 – Right_Internal_Ankle

19 – Right_Foot

20 – Right_External_Ankle

21 – Right_External_Knee

22 – Right_Pelvis

23 – Right_Axilla

24 – Right_Internal_Elbow

25 – Right_Internal_Wrist

26 – Right_External_Wrist

27 – Right_External_Elbow

28 – Right_Shoulder

We provide this dataset in two formats :

-Only annotations in csv text files. For this format, the images with the original sizes are used.

-Resized images and annotations with normalized shapes (64×128).

Annotations

Annotations are provided in the CSV format. The separator character is the ‘;’. The first column indicates the reference of the image followed by the coordinates of the landmarks in x and y.

Original images

This format uses the high resolution images of the Inria Person Dataset. Follow this link to download the original images. Annotations may be downloaded from HERE.

Two text files can be found, containing either 408 annotated shapes for training or 129 for testing (same categories as Inria Person Dataset).

Normalized images

These images and annotations are used in the evaluation of the ISPA paper. We define an area in the original images so that the Point Distribution Model after the resizing operation will have a maximal size in width and height of 64×128. Then, we increase this area to fill 50 pixels in the resized images as padding around the shape with a mirror method in each direction. It allows to take into account the area surrounding the person to compute local features etc. Finally, we resize the images octave by octave to avoid blur effect after brutal downsampling.

We also generated the flipped images named as *_a.txt. In total, this dataset contains 816 images and annotations for training and 258 images for testing.

Download the dataset with normalized images : TRAINING SET (62MB), TEST SET (16MB).

Triangulation

The joint triangulation may be found HERE. Indices refer to those indicated above.

Citation

If you use this dataset, please cite our work :

@INPROCEEDINGS{7306027,
author={O. Huynh and B. Stanciulescu},
booktitle={Image and Signal Processing and Analysis (ISPA), 2015 9th International Symposium on},
title={Clustering and classifying deformations for Shape Regression applied to the human body},
year={2015},
pages={25-30},
month={Sept},}

Disclaimer

THIS DATA SET IS PROVIDED “AS IS” AND WITHOUT ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, WITHOUT LIMITATION, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.

The images come from the Inria Person Dataset and may contain copyright issues. We take no guarantees or responsibilities arising out of any copyright issue. Use at your own risk.

If you have any questions, contact the authors of the paper Olivier Huynh: olivier[dot]huynh[at]mines-paristech[dot]fr or Bogdan Stanciulescu : bogdan[dot]stanciulescu[at]mines-paristech[dot]fr.