search

ina-foss / inaFaceAnalyzer

INA's library with pretrained models for gender and age prediction from faces.
Other
19 stars 9 forks source link
age age-bias age-estimation age-gender-estimation ass-subtitles deeplearning digital-humanities face face-analysis face-classification face-detection face-tracking gender gender-bias gender-classification gender-classifier pretrained-models python racial-bias video

readme

inaFaceAnalyzer: a Python toolbox for large-scale face-based description of gender representation in media with limited gender, racial and age biases

test py 3.7 test py 3.8 test py 3.9

PyPI version Docker Pulls status Documentation Status

About

inaFaceAnalyzer is a Python toolbox designed for large-scale analysis of faces in image or video streams. It provides a modular processing pipeline allowing to predict age and gender from faces. Results can be exported as tables, augmented video streams, or rich ASS subtitles. inaFaceAnalyzer is designed with speed in mind to perform large-scale media monitoring campaigns. The trained age and gender classification model provided is based on a ResNet50 architecture. Evaluation results are highly competitive with respect to the current state-of-the-art, and appear to reduce gender, age and racial biases.

Should you need further details regarding this work, please refer to the following paper:

@journal{doukhan2022joss,
  author = {David Doukhan and Thomas Petit},
  title = {inaFaceAnalyzer: a Python toolbox for large-scale face-based description of gender representation in media with limited gender, racial and age biases},
  journal = {JOSS - The journal of Open Source Software (currently being reviewed)},
  year = {submission in progress}
}

Have a look to sibling project inaSpeechSegmenter.

Statement of need

inaFaceAnalyzer is a Python framework aimed at extracting facial attribute information from massive video and image streams. This information can be used in a wide range of applications including biometrics, human-computer interaction, multimedia indexation, digital humanities and media monitoring.

inaFaceAnalyzer was realized to meet the needs of French National Audiovisual Institute (INA), in charge of archiving and providing access to more than 22 million hours of TV and radio programs. The emergence of computational digital humanities and data journalism has increased the need of INA's users to access meta-data obtained from automatic information extraction methods.

Since 2018, INA has realized several large-scale studies (up to 1 million hours of program analyzed) in the context of Gender Equality Monitor project, which aims at describing men and women representation differences in media based on speech time, TV text incrustations or facial attributes.

Since 2022, inaFaceAnalyzer is used in ARCOM's (French Regulatory Authority for Audiovisual and Digital Communication) annual report on Women representation in TV and radio. Channels statements are described jointly with inaFaceAnalyzer's automatic facial attribute estimates and inaSpeechSegmenter's speech-time estimates.

With respect to the high social impact associated to the studies using this software, inaFaceAnalyzer provide high accuracy prediction models. Being aimed at describing the representation of under-represented categories of people in media, it should minimize gender, age or racial biases that are known to also affect machine learning datasets and softwares. It is highly configurable, allowing to define trade-offs between accuracy and processing time depending on the scale of the analyses to be performed and on the available computational resources.

Installation

inaFaceAnalyzer requires Python version between 3.7 and 3.9. Python 3.10 is not yet supported due to onnxruntime-gpu dependency.

Installing from sources

apt-get install cmake ffmpeg libgl1-mesa-glx
git clone https://github.com/ina-foss/inaFaceAnalyzer.git
cd inaFaceAnalyzer
pip install .
./test_inaFaceAnalyzer.py # to check that the installation is ok

Installing from pypi on ubuntu

# for GPU support, cuda, cudnn and nvidia drivers should be already installed
apt-get install cmake ffmpeg libgl1-mesa-glx
pip install inaFaceAnalyzer

Using docker image

# download latest docker image from dockerhub
docker pull inafoss/inafaceanalyzer
# run docker image. setting --gpu argument allows to take advantage of
# GPU acceleration (non mandatory)
docker run -it --gpus=all inafoss/inafaceanalyzer /bin/bash
# lauch unit tests (non mandatory but recommended)
python test_inaFaceAnalyzer.py
# use any program or API
ina_face_analyzer.py -h

Using inaFaceAnalyzer command line programs

Several scripts are provided with the distribution:

A detailed listing of all the options from the command line programs is available using the -h argument. We guess you don't want to read the whole listing at this point, but you can have a look at it later 😉.

Displaying detailed manual

ina_face_analyzer.py -h

Process all frames from a list of video (without tracking)

Video processing uses the video engine and requires a list of input video paths, together with a directory used to store results in CSV. Program initialization time requires several seconds, and we recommend using large list of files instead of calling the program for each file to process.

# directory storing result must exist
mkdir my_output_directory
# -i is followed by the list of video to analyze, and -o is followed by the name of the output_directory
ina_face_analyzer.py --engine video -i ./media/pexels-artem-podrez-5725953.mp4 -o ./my_output_directory
# displaying the first 2 lines of the resulting CSV
head -n 2 ./my_output_directory/pexels-artem-podrez-5725953.csv
>> frame,bbox,detect_conf,sex_decfunc,age_decfunc,sex_label,age_label
>> 0,"(945, -17, 1139, 177)",0.999998927116394,8.408014,3.9126961,m,34.12696123123169
# using remote urls is also an option
ina_face_analyzer.py --engine video -i 'https://github.com/ina-foss/inaFaceAnalyzer/raw/master/media/pexels-artem-podrez-5725953.mp4' -o ./my_output_directory

The resulting CSV contains several columns:

Faster processing of a video

If computation time is an issue, we recommend using --fps 1 which will process a single frame per second, instead of the whole amount of video frames. When using GPU architectures, we also recommend setting large batch_size values.

# here we process a single frame per second, which is 25/30 faster than processing the whole video
ina_face_analyzer.py --engine video --fps 1 --batch_size 128 -i ./media/pexels-artem-podrez-5725953.mp4 -o ./my_output_directory

Using Tracking

Tracking allows to lower computation time, since it is less costly than a face detection procedure. It also allows to smooth prediction results associated to a tracked face and obtain more robust estimates. It is activated with the videotracking engine and requires to define detect_period, the time period (in frames) at which the face detector will be applied.

# Process 5 frames per second, use face detection for 1/3 and face tracking for 2/3 frames
ina_face_analyzer.py --engine videotracking --fps 5 --detect_period 3 -i ./media/pexels-artem-podrez-5725953.mp4 -o ./my_output_directory
# displaying the first 2 lines of the resulting CSV
head -n 2 ./my_output_directory/pexels-artem-podrez-5725953.csv
>> frame,bbox,face_id,detect_conf,track_conf,sex_decfunc,age_decfunc,sex_label,age_label,sex_decfunc_avg,age_decfunc_avg,sex_label_avg,age_label_avg
>> 0,"(945, -17, 1139, 177)",0,0.999998927116394,,8.408026,3.9126964,m,34.12696361541748,8.391026,3.8831162,m,33.831162452697754

Resulting CSV will contain additional columns with _avg suffixes, corresponding to the smoothed estimates obtained for each tracked face. It will also contain a face_id with a numeric identifier associated to each tracked face.

Exporting results

Results visualization allows to validate if a given processing pipeline is suited to a specific material. --mp4_export generate a new video with embedded bounding boxes and classification information. --ass_subtitle_export generate a ASS subtitle file allowing to display bounding boxes and classification results in vlc or ELAN, and which is more convenient to share..

# Process 10 frames per second, use face detection for 1/2 and face tracking for 1/2 frames
# results are exported to a newly generated MP4 video and ASS subtitle
ina_face_analyzer.py --engine videotracking --fps 10 --detect_period 2 --mp4_export --ass_subtitle_export  -i ./media/pexels-artem-podrez-5725953.mp4 -o ./my_output_directory
# display the resulting video
vlc ./my_output_directory/pexels-artem-podrez-5725953.mp4
# display the original video with the resulting subtitle files
vlc media/pexels-artem-podrez-5725953.mp4 --sub-file my_output_directory/pexels-artem-podrez-5725953.ass

Processing list of images

The processing of a list of images requires to use the image engine. A single resulting CSV will be generated with entries for each detected faces, together with a reference to their original filename.

# process all images stored in directory media, outputs a single csv file
ina_face_analyzer.py --engine image -i media/* -o ./myresults.csv
# display first 2 lines of the result file
head -n 2 myresults.csv
>> frame,bbox,detect_conf,sex_decfunc,age_decfunc,sex_label,age_label
>> media/1546923312_7cc94957e8_o.jpg,"(57, 104, 435, 483)",1.0,14.436495,3.5770981,m,30.770981311798096

Distributing analyses over a network

We provide two scripts allowing to perform distributed large-scale analyses.

ina_face_analyzer_distributed_server.py is in charge of distributing a list of documents to analyze to workers distributed over the network, and to define analysis options (fps, tracking, etc..). The server requires 2 positional arguments: its host name (or IP) and the path to a CSV containing one line per file to process together with the destination path of the results. Workers need to have writing permissions in the destination paths (mounted with NFS, sshfs, ...). Output directories are created on the fly if they don't exist. Jobs order is randomized before being distributed to the workers. When a destination file already exists, the corresponding analysis is skipped.

# a sample job list csv with 2 records and 4 columns
# source_path (mandatory input file path or url)
# dest_csv (mandatory output csv)
# dest_ass: to be used for exporting results to ass subtitles
# dest_mp4: to be used for exporting incrusted MP4 video
cat test.csv
>> source_path,dest_csv,dest_ass,dest_mp4
>> /home/ddoukhan/git_repos/inaFaceAnalyzer/media/pexels-artem-podrez-5725953.mp4,/tmp/csv/test1.csv,/tmp/ass/test1.ass,/tmp/mp4/test1.mp4
>> https://github.com/ina-foss/inaFaceAnalyzer/raw/master/media/pexels-artem-podrez-5725953.mp4,/tmp/csv/test2.csv,,
# the server define an analysis procedure at 1 FPS
# after initialization, it display a network adress to be passed to the workers
ina_face_analyzer_distributed_server.py blahtop.ina.fr test.csv --engine video --fps 1
>> parsing joblist test.csv
>> Total number of files to process: 2
Provide the following objet URI to remote ina_face_analyzer_distributed_workers:  PYRO:obj_4c027f06be5b40e7bcf2f3f1e235b68c@blahtop.ina.fr:33825

ina_face_analyzer_distributed_worker.py is in charge of computing analyses and writing results to a centralized storage directory. It requires the network adress displayed by the server in order to communicate. A good practice is to launch one worker per available GPU and set CUDA_AVAILABLE_DEVICES. Several workers can process the list of the server in parallel.

# CUDA_AVAILABLE_DEVICES=2 is non mandatory and tells the worker
# to use a single GPU with id 2.
# the PYRO:obj_ adress is displayed when lauching the server and
# should copy/pasted when launching the worker
CUDA_AVAILABLE_DEVICES=2 ina_face_analyzer_distributed_worker.py PYRO:obj_4c027f06be5b40e7bcf2f3f1e235b68c@blahtop.ina.fr:33825
>> received job https://github.com/ina-foss/inaFaceAnalyzer/raw/master/media/pexels-artem-podrez-5725953.mp4 /tmp/test2.csv nan nan
>> received job /home/ddoukhan/git_repos/inaFaceAnalyzer/media/pexels-artem-podrez-5725953.mp4 /tmp/test1.csv /tmp/test1.ass /tmp/test1.mp4
>>all jobs are done

Using inaFaceAnalyzer API

inaFaceAnalyzer's API documentation is available on readthedocs.io.

Several tutorial notebooks stored in directory tutorial_API_notebooks:

We provide below a brief description of these notebooks, together with links allowing to run them remotely in Google's colab platform. When using colab platform, we recommend to take advantage of GPU acceleration.

Contributing

Please feel free to open issues if you have any questions or suggestions, or if you want to contribute to inaFaceAnalyzer development. Every contribution is very welcome! Please read CONTRIBUTING.md for more details.

CREDITS

This work has been partially funded by the French National Research Agency (project GEM : Gender Equality Monitor : ANR-19-CE38-0012) and by European Union's Horizon 2020 research and innovation programme (project MeMAD : H2020 grant agreement No 780069).

We acknowledge contributions from Zohra Rezgui who trained first models and wrote the first piece of code that lead to inaFaceAnalyzer during her internship at INA.

@techreport{rezgui2019carthage,
  type = {Msc. Thesis},
  author = {Zohra Rezgui},
  title = {Détection et classification de visages pour la description de l’égalité femme-homme dans les archives télévisuelles},
  submissiondate = {2019/11/19},
  year = {2019},
  url = {https://www.researchgate.net/publication/337635267_Rapport_de_stage_Detection_et_classification_de_visages_pour_la_description_de_l'egalite_femme-homme_dans_les_archives_televisuelles},
  institution = {Higher School of Statistics and Information Analysis, University of Carthage}
}