June 26, 2020

Everything you need to know about Custom Facial Recognition 3.1

Table of Contents:

Unlock the power of Custom Facial Recognition (CFR) with Clarifai's all new CFR 3.1.  CFR 3.1 delivers major advancements across all categories of facial recognition performance, with overall performance boosts of  20-25% on our evaluations. Now you can build your AI solution with our most advanced facial recognition technology yet. Read about how we did it.

Innovations

We are really excited about the impact that applied research can have in AI. We have made significant advancements in our facial recognition technology as we work to provide ultra-efficient solutions for a changing world. People tend to be the most important objects in our photos and videos, so it comes as no surprise that fast, accurate facial recognition technologies are powering solutions in retail, travel, tourism, public safety and more. 

More, better training data 

Training data is key to any AI model, and we trained or new CFR model with millions of additional images and 7x more unique identities than our previous model. 

cfr31

3D Landmark Detection

We have added entirely new techniques to our tool box with 3D landmark detection and alignment. 3D landmark detection allows our model to identify the location of specific facial features, like the corners of eyes and mouths, and then use these anchor points to align faces in a consistent way. This means that users of our model can account for the direction a person is facing when analyzing an image, which helps you tailor your applications for more accurate results.  

3d_landmark

Model Architecture

We have built our CFR model with an entirely new model architecture based on cutting edge research in facial recognition. Our new architecture has 20x the number of layers of our previous model, while outputting a feature representation that requires half the memory. This makes for a responsive model, with a better understanding of the human face.

layers

New algorithms 

Advanced new algorithms improve the discriminative power of our new model without requiring a lot of additional computational power. This means that our model is better at distinguishing between similar faces and it runs fast. 

angular_margin_loss-2

Performance improvements

In internal evaluations, CFR 3.1 significantly outperformed open source implementation of the ArcFace model, as well as our current (beta) production CFR model in both search accuracy and face verification tasks. 

Search Accuracy

Our customers love how visual search can help them quickly find images of specific people within their dataset. Without even training a custom model, you can use AI to find people with the click of a button.

CFR 3.1 outperformed both of our benchmark datasets across all categories that we tested, with accuracy boosts of up to 30% over our previous model. 

search_accuracy

Face Verification 

For verification technologies, the "true positive rate" (TPR) shows the percentage of cases where the model correctly identifies a match between a pair of face images belonging to the same identity.

Once again, CFR 3.1 outperformed all other models, with TPR performance enhancements of over 40% as compared to our previous model. 

sarah_verifier

How to use CFR 3.1

All new facial detection apps on the Clarifai platform use CFR 3.1, so there is nothing special for you to do to take advantage of this new technology. To get started, just choose "Face" as your base workflow when you create a new model. You can also add face detection to an existing workflow

Plus, apps built with pervious versions of our CFR models will continue to be supported.

blog-portal-signup-cta