Table of Contents
- Face detections
- Model Validation
- Rendering the paper
The goal of this project was to analyze social information accessible to young children through headcam videos. The raw headcam data consists of 36 videos, one per subject, with the subjects belonging to 3 age groups: 8, 12, and 16 months (with equal subjects per group). The computational pipeline consists of first detecting faces in videos (where all video processing is done on Sherlock, an HPC cluster available to Stanford faculty) and subsequently analyzing face detections along with posture and orientation annotations and demographics.
All face detections were done on Sherlock, using SLURM for job management.
First steps involve moving the 36 xs-face videos to the computing environment of choice, in our case Sherlock. To do this we used Globus. In Sherlock, the videos are located in the directory $PI_HOME/headcam_videos_original. Since the videos were recorded upside-down, we rotated all of them using scripts in preprocessing. We then extracted frames using ffmpeg v3.5.1, which along with their rotated versions are located in $PI_HOME/frames and $PI_HOME/rotated_frames, respectively.
Face, Hand, Pose detection
Once all the frames were extracted, we ran the MTCNN algorithm using the mtnn scripts. We used a Tensorflow implementation of the MTCNN algorithm found here. The script looks through every subject folder (i.e. "XS_0803") in the $PI_HOME/frames_rotated folder and runs the face detector (in a CPU env), which outputs a csv corresponding to that subject in a tmp folder. The csv contains the bounding box coordinates if a face was detected in the frame. Once all jobs finished, we combined the csvs in tmp to form a consolidated csv that was then stored here.
Unlike MTCNN, OpenPose did not work well on folders of frames (averaging about 28K per subject), but instead had its own utility for extracting frames and rotating them from videos, so that's what we used. In contrast to MTCNN which was run on the native Sherlock environment, we ran OpenPose through a docker container that already had the pre-compiled OpenPose binary. The Docker container with instructions on how to run it on Sherlock can be found here. This binary must be run in a GPU environment. The output from OpenPose is stored in $PI_HOME/openpose_json_output, with a json file corresponding to every frame. The json stored the locations of 18 different body parts, 69 detailed face points, and 42 detailed hand points. In order to consolidate the json into one csv, we ran a script inside the openpose folder and stored the output in final_output. However, this file was too large to save in git, so we instead uploaded a version to google drive.
In order to validate the models, we selected a sample of gold set images to annotate inside ground_truth_creation. These candidates were then downloaded from Sherlock onto my local machine, where they were annotated with a tool called Rectlabel. This tool created an annotation folder in every subject folder, containing all bounding box coordinates from the squares we drew in XML. The script inside ground_truth_creation essentially extracted the face info from this setup and created a ground truth file. All analyses comparing this ground truth with actual results are detailed in the .Rmd file in writeup-cogsci.
In order to not bias the evaluation because of the relatively rare appearance of faces in the dataset, two gold samples were selected: one high density sample consisting of a large number of face detections, and one random sample from the remaining frames.
The final face detection results are stored in final_output. These data were merged with the corresponding posture, orientation, and demographics csvs to form one large consolidated csv that is not in the repistory as it is too large for GitHub (>100MB).