This repository contains the relevant data, evaluation and result files we used for the results shown in our paper titled: _Automatic infant motion extraction from videos: comparing six deep neural network 2D pose estimation methods_ [ADD LINK when paper available online], except for the datasets.
Our own dataset of infant recordings is sensitive and not anonymized data, and cannot be shared publicly.
The MINI-RGBD dataset of synthetic infants is available at its official source: [https://www.iosb.fraunhofer.de/en/competences/image-exploitation/object-recognition/sensor-networks/motion-analysis.html#mini-rgbd]
## Data & Results
The sub component with this name contains all the data and the analysis programs we used to generate our results.
## Docker containers for pose estimation methods
One of the goals of the paper was to make pose estimation methods more easily accessible to researchers that are not in computer vision or even computer science in general, and in particular for developmental psychology researchers.
The Docker containers themselves are available from our DockerHub page [https://hub.docker.com/u/humanoidsctu], and only need to be downloaded and used by following the instructions.
In this osf repository, everything needed to build the Docker containers we used are available in the _Docker_ sub-component, including any file we modified and the weights of the networks in case they become unavailable from their original source.
DeepLabCut's official github repository already contains easy instructions to create a conda environment and install it, so no Docker was needed.
However, it still has its own folder with some help and instructions, the model's weights, and a slightly modified main script.
