This module is still under development. Code is subject to major changes.
Detect and redact faces in images and videos. Save the results to JPG, PNG, MP4, or as JSON data files.
First, source the filepaths environment variables
# source environment variables used for examples
source data/examples/filepaths.sh
Simple face detection and blurring for images
vf pipe open -i $FACE_IMAGE detect -m yoloface redact draw display
Simple face detection and blurring for videos
# Detect and blur faces in a video
vf pipe open -i $FACE_VIDEO detect -m yoloface redact draw display --auto
Save images or video
# Save images
vf pipe open -i $FACE_IMAGE detect -m yoloface blur save-images -o $DIR_IMAGES_OUT
# Save video
vf pipe open -i $FACE_VIDEO detect -m yoloface redact draw save-video -o $DIR_VIDEO_OUT
Blur the detect face regions
# Detect and blur faces and save to new file
vf pipe open -i $FACE_IMAGE detect -m yoloface blur save-images -o $DIR_IMAGES_OUT
# Detect and blur faces and save to new file, draw face bbox
vf pipe open -i $FACE_IMAGE detect -m yoloface redact draw save-images -o $DIR_IMAGES_OUT
# rewrite as multi-line command for clarity
vf pipe \
open -i $FACE_IMAGE \
detect -m yoloface \
redact \
draw \
save-images -o $DIR_IMAGES_OUT
Add save-images or save-video to the pipe commands to save redacted frames
# Save blurred face image
vf pipe open -i $FACE_IMAGE detect -m yoloface draw save-images $DIR_IMAGES_OUT
# Save blurred face video
vf pipe open -i $FACE_VIDEO detect -m yoloface redact draw save-video
Blur Directory of Images or Videos
# Directory of JPG images
vf pipe \
open -i ../data/images/ --exts jpg \
detect -m yoloface \
redact \
save-image -o ~/Downloads/ --suffix _redacted
# Directory of MP4 videos
vf pipe
open -i ../data/images/ --exts mp4 \
detect -m yoloface \
redact \
save-video -o ~/Downloads/ --suffix _redacted
Expand BBox (eg: to cover ears)
# Expand the bounding box to blur more than the face
vf pipe \
open -i ../data/media/input/samples/faces.jpg \
detect -m yoloface \
blur --expand 0.5 \
display
Blur all faces in a video and export detections to JSON file
# Blur faces in a single image
vf pipe \
open -i ../data/media/input/samples/faces.mp4 \
detect -m yoloface \
redact \
save_data -o ../data/media/output/
Blur faces and save detections to JSON
# Blur faces in a single image
vf \
pipe open -i ../data/media/input/samples/faces.mp4 \
detect -m yoloface \
redact \
save_data -o ../data/media/output/faces.json
Detector ensemble
# Merge detections from multiple models
vf pipe \
open -i ../data/media/input/samples/faces.jpg \
detect -m yoloface \
detect -m yoloface \
merge --to face \
blur -n face \
draw -c 255 255 255 -n face \
display
Multi-detector ensemble with post-processing NMS
#!/bin/bash
# save this to detect.sh and run "bash detect.sh"
FP_VIDEOS=/path/to/videos
FP_DETECTIONS=/path/to/detections
GPU=0 # choose GPU index
export CUDA_VISIBLE_DEVICES=${GPU}; vf pipe \
open -i ${FP_VIDEOS} -e mp4 --slice 0 1 \
resize -w 960 -f original \
detect -m retinaface -n retinaface_0 \
detect -m retinaface -r 90 -n retinaface_90 \
detect -m retinaface -r 270 -n retinaface_270 \
merge --to retinaface \
save_data -o ${FP_DETECTIONS}/retinaface.json
export CUDA_VISIBLE_DEVICES=${GPU}; vf pipe \
open -i ${FP_VIDEOS} -e mp4 --slice 0 1 \
resize -w 960 -f original \
detect -m yoloface -n yoloface_0 \
detect -m yoloface -r 90 -n yoloface_90 \
detect -m yoloface -r 270 -n yoloface_270 \
merge --to yoloface \
save_data -o ${FP_DETECTIONS}/yoloface.json
export CUDA_VISIBLE_DEVICES=${GPU}; vf pipe \
open -i ${FP_VIDEOS} -e mp4 --slice 0 1 \
resize -w 960 -f original \
detect -m yoloface -n yoloface_0 \
detect -m yoloface -r 90 -n yoloface_90 \
detect -m yoloface -r 270 -n yoloface_270 \
merge --to yoloface \
save_data -o ${FP_DETECTIONS}/yoloface.json
If GPU RAM is limited, run each command separately
#!/bin/bash
# save this to detect.sh and run "bash detect.sh"
FP_VIDEOS=/path/to/videos
FP_METADATA=/path/to/metadata
DETECTORS=(yoloface retinaface yoloface)
ROTS=(0 90 270)
for DETECTOR in ${DETECTORS}; do
for ROT in ${ROTS}; do
vf pipe \
open -i ${FP_VIDEO} -e mp4 \
resize -w 960 -f original \
detect -m ${DETECTOR} -n ${DETECTOR}_${ROT} -t 0.6 \
save_data -o ${FP_METADATA}/${DETECTOR}_${ROT}.json
done
done
On a remote computer create a tmux session to monitor and detach
tmux new -s detect
conda activate vframe
# run above script
# detach: ctl+b then d
# reattach: tmux a -t detect
Then merge all JSON files
vf dev merge-detections \
-i [json1] \
-i [json2] \
-i [json3] \
-o ${FP_METADATA}/merged.json
Then run NMS on the merged detections
vf dev merge-detections-nms \
-i ${FP_METADATA}/merged.json \
-o ${FP_METADATA}/merged_nms.json
Then blur the videos uses pre-computed detections
FP_OUT=/path/to/videos_blurred/
vf pipe \
open -i ${FP}/merged.json \
redact \
draw -C 255 255 255 \
save-video -o ${FP}
To be implemented:
- reattach audio track
Detecting Small Faces
By default, the detectors use the image size settings in the ModelZoo YAML configuration files. These are general settings and should be tailored to your media. Supply a --width or --height argument to increase or decrease the input size. Increasing the input size helps detect smaller faces but is more computationally intensive. Decreasing the input size detects less smaller faces, but is faster.
# Increase size (slows down, but detects more small faces)
vf pipe
open -i ../data/media/input/samples/faces.jpg \
detect -m yoloface --width 960 \
redact \
display
# Decrease size (speeds up, but detects less small faces)
vf pipe \
open -i ../data/media/input/samples/faces.jpg \
detect -m yoloface --width 448 \
redact \
display
Development of the VFRAME face blurring tools was supported by NL Net Privacy Enhancing Technologies during 2019 - 2020.
Research and development: Adam Harvey, VFRAME