6.6 KiB

Raw Blame History

OpenFace

This is the documentation of OpenFace.

The code and issue tracker is available on GitHub at cmusatyalab/openface.

Please join the cmu-openface group or the gitter chat for discussions and installation issues.

Development discussions and bugs reports are on the issue tracker.

This is a Python and Torch implementation of the CVPR 2015 paper FaceNet: A Unified Embedding for Face Recognition and Clustering by Florian Schroff, Dmitry Kalenichenko, and James Philbin at Google using publicly available libraries and datasets. Torch allows the network to be executed on a CPU or with CUDA.

Crafted by Brandon Amos in the Elijah research group at Carnegie Mellon University.

Isn't face recognition a solved problem?

No! Accuracies from research papers have just begun to surpass human accuracies on some benchmarks. The accuracies of open source face recognition systems lag behind the state-of-the-art. See our accuracy comparisons on the famous LFW benchmark below.

Please use responsibly!

We do not support the use of this project in applications that violate privacy and security. We are using this to help cognitively impaired users to sense and understand the world around them.

Overview

The following overview shows the workflow for a single input image of Sylvestor Stallone from the publicly available LFW dataset.

Detect faces with a pre-trained models from dlib or OpenCV.
Transform the face for the neural network. This repository uses dlib's real-time pose estimation with OpenCV's affine transformation to try to make the eyes and nose appear in the same location on each image.
Use a deep neural network to represent (or embed) the face on a 128-dimensional unit hypersphere. The embedding is a generic representation for anybody's face. Unlike other face representations, this embedding has the nice property that a larger distance between two face embeddings means that the faces are likely not of the same person. This property makes clustering, similarity detection, and classification tasks easier than other face recognition techniques where the Euclidean distance between features is not meaningful.
Apply your favorite clustering or classification techniques to the features to complete your recognition task. See below for our examples for classification and similarity detection, including an online web demo.

What's in this repository?

batch-represent: Generate representations from a batch of images, stored in a directory by names.
demos/web: Real-time web demo.
demos/compare.py: Demo to compare two images.
demos/vis-outputs.lua: Demo to visualize the network's outputs.
demos/classifier.py: Demo to train and use classifiers.
evaluation: LFW accuracy evaluation scripts.
openface: Python library code.
images: Images used in the README.
models: Model directory for openface and 3rd party libraries.
training: Scripts to train new OpenFace models.
util: Utility scripts.

Citations

Please cite this repository if you use this in academic works.

@misc{amos2015openface,
    author       = {Amos, Brandon and Harkes, Jan and Pillai, Padmanabhan and Elgazzar, Khalid and Satyanarayanan, Mahadev},
    title        = {OpenFace 0.1.1: Face recognition with Google's FaceNet deep neural network},
    month        = oct,
    year         = 2015,
    doi          = {10.5281/zenodo.32148},
    url          = {http://dx.doi.org/10.5281/zenodo.32148}
}

Acknowledgements

The fantastic Torch ecosystem and community.
Alfredo Canziani's implementation of FaceNet's loss function in torch-TripletEmbedding.
Nicholas Léonard for quickly merging my pull requests to nicholas-leonard/dpnn modifying the inception layer.
Francisco Massa and Andrej Karpathy for quickly releasing nn.Normalize after I expressed interest in using it.
Soumith Chintala for help with the fbcunn example code.
NVIDIA's academic hardware grant program for providing the Tesla K40 used to train the model.
Davis King's dlib library for face detection and alignment.
Zhuo Chen, Kiryong Ha, Wenlu Hu, Rahul Sukthankar, and Junjue Wang for insightful discussions.

Licensing

The source code and trained models nn4.v1.t7 and celeb-classifier.nn4.v1.t7 are copyright Carnegie Mellon University and licensed under the Apache 2.0 License. Portions from the following third party sources have been modified and are included in this repository. These portions are noted in the source files and are copyright their respective authors with the licenses listed.

Project	Modified	License
Atcold/torch-TripletEmbedding	No	MIT
facebook/fbnn	Yes	BSD

6.6 KiB Raw Blame History