A 3-d human skeleton pose annotation dataset released by Max Planck Institute for Informatics (MPII).
Download the dataset zip package (mpi_inf_3dhp.zip) from here.
Unzip, then read the mpi_inf_3dhp/README.txt file.
ATTENTION:
You would need to read and review the configuration under conf.ig before you can proceed with downloading the dataset !!!
Find and set the following configurations in conf.ig:
subjects=(1 2 3 4 5 6 7 8)
ready_to_download=1Use the script get_dataset.sh to download the training set and get_testset.sh for the test set. Make sure you have approx 25GB space in this path to download the complete training set. The test set needs another 7GB and can be downloaded with get_testset.sh.
Downloading takes a long time.
Optional:
Use the following command to run the shell script in the background:
nohup sh get_dataset.sh &The image frames of the dataset are given in the form of video sequences. Use the script get_dataset_img.sh provided in this repository to extract frame images from video sequences.
Be cautious that the generated images will consume super huge storage space!!!
Note:
The following command
ffmpeg (ffmpeg -i "<some_folder>/video_X.avi" -qscale:v 1 "<some_folder>/img_X_%06d.jpg")generates image frames with valid correspondence to the annotations.
About ffmpeg:
ffmpeg -i *.avi -vf "select=between(n\,84\,208)*not(mod(n\,25))" -vsync 0 ./images/image_%06d.jpg -vf: select filter, between(n,*) means split from 84 frame to 208 frame.
not(mode(n\, K)): output 1 image from every 25 frames.
ffmpeg -i *.avi -r 1 -vf fps=fps=1 ./images/image_%06d.jpg -vf fps=fps=1(-r 1): the rate of screenshot is 1 frame per second.
To compress the frame images, run scale_img.py.
To extract 2D and 3D annotation data of joint positions from .mat files to hdf5 format, run gen_h5.m.
Edit configuration of this MATLAB script by setting first two lines before running it.
Run gen_file_list.py to generate train.txt and test.txt containing available image paths.