RedHen Gestures

This is the repository for all the code that was ceated during the GSoC 2021 at Red Hen Lab for the project topic "Red Hen Open Dataset for gestures with performance baselines at Red Hen Lab".

For more info on the work and some information / How to use / documentations on Singularity and HPC, check my blog @ swadesh13.github.io/gestures-dataset-blog

The OpenPose container used in this project is derived from frankierr/openpose_containers:focal_nvcaffe at Docker. (Check the singularity.df file for reference). workflow.py file works as the entry to all the code and handles all the tasks.

The src directory contains all the code. There is a code directory for python files such as workflow.py. Also, the singularity def file and .openpose_env are under the singularity directory. I updated the .openpose_env and the def files, so I have kept them here.

Steps followed after screating a singularity sandbox from openpose.def file

• Changed access permissions of /.openpose_env to allow execution.
• Changed the /.openpose_env to add the workflow.py file path.
• Openpose searches for models directory at OPENPOSE_SRC i.e. base OpenPose folder. So, copied the directory /home/opt/openpose_models to /home/opt/openpose/models
• Added this github repo to /home/opt/openpose/.

Structure of the Code

Go to src/code. Workflow.py handles all input and outputs. Inside data, data related tasks such as arranging keypoints to numpy data is present in gestures_data.py. The model directory contains the training and detection code. The pose folder contains code to generate keypoints from OpenPose output in the keypoints.py file. utils contains some simple utility stuff. config.py includes the basic default configuration. singularity. singularity folder contains the basic def file on which the container is built and also the .openpose_env file, which contains an extra environment path to be added. models contains the model with default parameters.