HUMAN POSE UTIL

Various utilities for human pose estimation.

Setup

Clone the repository and install using pip install -e.

cd path/to/parent_folder
git clone https://github.com/jackd/human_pose_util.git
pip install -e human_pose_util

Datasets

This repository comes with support for Human3.6M (h3m) and HumanEva_I (eva) datasets. Due to licensing issues these are not provided here - see the respective websites for details.

Setting up datasets

Human3.6M (h3m)

To work with the Human3.6M dataset, you must have the relevant .cdf files in an uncompressed local directory, referenced here as MY_H3M_DIRECTORY. For licensing reasons, we cannot provide the raw Human3.6m data. Please consult the website to source the original data. This directory must have the following structure:

- MY_H3M_DIRECTORY
  - D2_positions
    - S1
      - Directions.54138969.cdf
      - ...
    - S5
      - ...
    ...
  - D3_positions
    - S1
    ...
  - D3_positions_mono
    - S1
    ...
  - Videos
    - S1
    ...

Videos aren't used in module, though the dataset has a video_path attribute which assumes the above structure.

To let the scripts know where to find the data, run the following in a terminal

export H3M_PATH=/path/to/MY_H3M_DIRECTORY

Consider adding this line to your .bashrc if you will be using this a lot.

HumanEva_I (eva)

To work with the HumanEva_I dataset, you must have the uncompressed data available in MY_EVA_1_DIR which should have the following structure:

- MY_EVA_1_DIR
  - S1
    - Calibration_Data
      - BW1.cal
      ...
    - Image_Data
      - Box_1_(BW2).avi
      ...
    - Mocap_Data
      - Box_1.c3d
      - Box_1.mat
      ...
    - Sync_Data
      - Box_1_(BW1).ofs
      ...
  - S2
    ...
  ...

Image_Data is not used in this module, thought the dataset has a video_path attribute which assumes the above structure.

To let scripts know where to find the data, run the following in a terminal

export H3M_PATH=/home/jackd/Development/datasets/human3p6m/data

Consider adding this line to your .bashrc if you will be using this a lot.

Registering a new dataset

A new dataset can be registered using

human_pose_util.register.dataset_register[dataset_id] = {
    'train': train_datastet,
    'eval': eval_dataset,
}

If your dataset uses a different skeleton from those provided (see human_pose_util.skeleton.Skeleton), you'll need to precede this with a similar skeleton registration line

human_pose_util.register.skeleton_register[my_skeleton_id] = my_skeleton

After that, training/inference can procede as normal.

See human_pose_util.dataset.h3m and human_pose_util.dataset.eva for examples.

Visualizations

Visualizations are dependent on matplotlib or glumpy, though these are not included in the dependencies.

Installing matlplotlib:

pip install matplotlib

Installing glumpy:

git clone https://github.com/glumpy/glumpy.git
cd glumpy
python setup.py install

TODO

• Continue removing all dataset.interface stuff. Mostly there with h3m, but start with dataset.h3m.pose_sequence. It's overly complicated - just map lists of dictionaries (and use hdf5 instead of dicts if memory becomes too intensive - unlikely for just pose stuff, though maybe necessary for images/heatmaps).
• dataset/mpi_inf/README.md TODOs