Release notes

• New face recognition model

Introducing a smaller, faster, and more accurate face recognition model.

• The new and improved face detection model

Introducing improved face detection model, more robust to various challenging conditions such as faces with high variability in scale, illumination, pose and occlusion.

• New fast mode of VNN tracking algorithm

Introducing new VNN algorithm fast mode which significantly improves performance at the cost of feature points precision while keeping the same precision of head pose.

• Improved tracking performance of VNN algorithm

VNN tracking algorithm now works with higher FPS, with significant improvements on devices such as high-end mobile and desktop devices.

• Reduced tracking noise in VNN tracking algorithm on mobile platforms – Android, iOS

New VNN tracking models now work with less tracking jitter.

Known issues

• The new improved face detection model

Introducing improved face detection model, more robust to various challenging conditions such as faces with high variability in scale, illumination, pose and occlusion.

• New fast mode of VNN tracking algorithm

Introducing new VNN algorithm fast mode which significantly improves performance at the cost of feature points precision while keeping the same precision of head pose.

• Improved tracking speed of VNN algorithm

VNN tracking algorithm now works with higher FPS, with significant improvements on devices such as high-end mobile and desktop devices

• Reduced tracking noise in VNN tracking algorithm (iOS, Android)

New VNN tracking models now work with less tracking jitter.

Known issues

• New runner on mobile platforms – Android and iOS

Introducing new runner – TensorFlow Lite, optimized for running neural networks and significantly improving the performance of visage|SDK algorithms.

• New Face recognition model

Introducing a smaller, faster, and more accurate face recognition model that completely replaces the old model that will no longer be distributed.

The model is available for OpenVino™ runner on desktop platforms and Tensorflow Lite runner on mobile platforms.

Known issues

• A novel experimental tracking algorithm – VNN – introduced

The new algorithm minimizes jitter, increases tracking accuracy and robustness but reduces tracking performance (speed). It is demonstrated in ShowcaseDemo and FaceTracker2 samples via Ultra tracking configuration.

• New neural network runner provided – Intel’s OpenVINO™

Significantly improves the performance of age estimation, face recognition and face tracking with VNN algorithm on Intel 64-bit processors.

• Ear tracking

Additional 24 feature points on ears are now tracked (12 points per ear). Ear tracking is configurable through the tracker configuration file or API.

• Iris tracking

Face data from tracker and detector now includes information about iris diameter.

• VisageConfiguration API introduced

It is now possible to modify specific tracker configuration parameters via an interface during tracking.

• Age estimation accuracy improved

• Ear tracking introduced: Tracker can track 24 additional feature points on ears (12 points per ear); Configurable via configuration file.
• Iris tracking: Face data from tracker and detector now includes information about iris diameter.
• Age estimation accuracy improved
• Tracking performance (speed) improved

• Improved smoothing filter
  Smoothing of feature points is performed using multiple filters. For still face, higher amount of smoothing is applied while fast movements are less smoothed in order to avoid noticeable delay. Increased stability of feature points and head position especially in profile and half-profile pose.
• Refactoring of frame preprocessing resulting in more stable FPS and improved accuracy on high-resolution frames
  The core tracking loop was reimplemented to make the tracking frame rate less dependent on the size of the face in the image. This fixes performance drops in cases where the face takes up a small portion of the frame. Additionally, noise introduced by resizing of higher-resolution frames is reduced which results in more stable tracking.

Platform: HTML

• API upgraded to use memory arrays
  API for fetching tracking data has been modified to return memory arrays. Improves performance and simplifies memory managment of tracked data.

Platforms: iOS, Android

• ShowcaseDemo introduces example of tracking from video including source code.