search

brialorelle / headcam-algo

Repository for development/validation of algorithms to extract automated information from headcam data
2 stars 1 forks source link

readme

A longitudinal analysis of the social information in infants' naturalistic visual experience using automated detections

Reproducing analysis in the submission

The main analyses and writeup for the paper are in the paper/preprint/manuscript_june2021_apa_blinded.Rmd. Data for the paper can be found at https://osf.io/cdhw4/ in data_preprocessed_2021.zip; this folder should be unzipped and placed in the main directory/data.

  1. OpenPose detections were obtained using the pipeline outlined below (openpose_detections/openpose_pipeline.ipynb).
  2. Bounding boxes were extracted using preprocessing/2_get_bounding_boxes_per_vid.py and preprocessing/step3_merge_per_vid_bounding_boxes.R
  3. Human annotations for faces/hands obtained using AMT; basic preprocessing is in preprocessing/step0_get_human_annotations/face_hands/analysis a
  4. Detections were preprocessed and joined with metadata to create intermediate data files, using preprocessing/step4_create_bounding_box_summaries.Rmd and preprocessing/step5_preprocess_fullset_from_bb_outputs.Rmd

Running OpenPose Detections

This repository contains code designed to run on infant headcam videos at scale, using the Openpose pose detector to extract information about facial and manual cues in the infants' environments.

Getting Started

This repository was made to run on the Sherlock computing cluster at Stanford. As such, some analysis steps rely on the Slurm Workload Manager to submit compute-intensive jobs. You may need to modify the job submission code if you are running on a different cluster.

Prerequisites

You should have conda and Singularity installed on your machine.

Installing

Here's what's needed to setup the project.

Clone the repo:

git clone https://github.com/brialorelle/headcam-algo.git
cd headcam-algo/

Next, make setup.sh executable and run it to setup the conda environment and pull the openpose docker into a location with at least 2GB of space:

chmod +x setup.sh
./setup.sh /path/to/desired/openpose_install_dir

Download Dataset

With the headcam conda environment active, run the download_databrary_volume.py script to download volume 564 on databrary, "Head cameras on children aged 6 months through 31 months", and then run clean_video_repo.sh to format the filenames properly and remove extraneous, non-video files:

You will need roughly 500GB of space to download this volume.

python misc/download_databrary_volume.py 564 -o /path/to/desired/video_download_dir
chmod +x ./misc/clean_video_repo.sh
./misc/clean_video_repo.sh /path/to/desired/video_download_dir

(In order to download a different Databrary dataset, simply specify the appropriate volume number for that dataset. clean_video_repo.sh is not intended for use on other datasets.)

Analysis

Follow the instructions in the README in the openpose_detections folder in order to run Openpose on all the videos in the dataset and save outputs in CSV format.