Face Recognition
About | FAQ | Contact | Blog | Press | Jobs | Research | Developers

Welcome to face.com research

Our primary mission at face.com is to research and develop the most accurate face recognition technology, aimed to tackle the challenging photos of the web. Working with photos on the web means working with huge amounts of data and also dealing with unconstrained conditions. Such photos are generaly taken in difficult settings, where the faces are not necessarily frontal, nor well lit and may be partly occluded or taken with low resolution cameras. Hence, unconstrained face recognition is very different from the conservative one which is generally employed in security facilities, border control and other controlled spaces. We have decided to promote the face recognition research community by sharing some of our research activities and tools in order to advance the field.

New: Improved Recognition Results

We are glad to announce a major improvement in our core technology, greatly boosting recognition performance. There's still fine tuning, tweaking and testing to be done, but we expect to make it available through the API soon. We'd like to share this with the research community first, presenting some preliminary metrics and visualizations.

  • Error rates have been reduced by approx. 30% in the average performance.
  • At zero false positives (errors), we have almost doubled the best recall rate to date. i.e. in selected benchmarks, requiring the new algorithm to never answer mistakenly, it will return twice the number of correct answers than it previously did.
  • We boast many more high-confidence recognitions than before - not just saying who it is, but saying it with high confidence
  • Overall more robustness to aging, pose, expression and illumination factors imposed in photos.
  • Improved performance for models that have been constructed from a single photo.

We've also employed face.com's next face recognition engine to the Labeled Faces in the Wild (LFW) benchmark which has become the de-facto standard testbed for unconstrained face recognition, without any dataset specific pre-tuning. The obtained mean accuracy is 91.3% +/- 0.3, achieved on the test set. Notably, much of the obtained improvement is achieved at the high-valued performance range of low false positive rate.

[See more]