v0.8.2.0 Release

Changelog (v0.8.2.0)

• Face Expression Estimation
  • 6DOF head pose now available.
    • Enable by setting NvAR_Parameter_Config(PoseMode) to 1. Default is 0.
    • Get the pose translation using NvAR_Parameter_Output(PoseTranslation).
    • Camera intrinsics can be set using NvAR_Parameter_Input(CameraIntrinsicParams).
  • Expression estimation model updated. The new model is more accurate, and enables subtle expressions such as asymmetric brows, inner/outer brow separation, asymmetric smile and frown, jaw left/right, squint shapes, crossing eye gaze, cheek puff, and more.
    • Enable cheek puff shapes by setting NvAR_Parameter_Config(EnableCheekPuff) to 1.
  • New face model for visualization recommended for FaceExpressions feature. The model has updated blendshapes, and face area partitioning.
    • Enable using --render_model=face_model3.nvf in ExpressionApp.
• Eye Contact
  • Performance improvements
    • CUDA graphs functionality to enable CUDA optimization now available for Eye contact feature.
    • Use the parameter NvAR_Parameter_Config(UseCudaGraph) to enable/disable CUDA graphs. Default is OFF

v0.8.1.0 Release

Changelog (v0.8.1.0)

• NEW! Eye Contact feature: an AI algorithm to help users keep their gaze engaged in video communication. The feature jointly estimates a user’s gaze direction and redirects it to frontal in video sequences.
• NEW! Face Expression Estimation (Beta) feature estimates facial expression coefficients From the video or the provided facial landmarks. ExpressionApp is added to demonstrate the new Face Expressions feature.
• NEW! Default face model for the Face 3D mesh and tracking feature, face_model2.nvf, now ships with the SDK. The old SFM based face_model0.nvf is no longer required.
• 3D Body Pose Estimation:
  • NEW! Added the support for Multi Person Tracking. This feature is supported by the Windows SDK only.
  • FocalLength is now a NvAR_Parameter_Input. Users can now change FocalLength at every NvAR_Run() without having to call NvAR_Load().
  • The reference pose returned by the feature has been updated
• Facial landmark estimation
  • NEW! There are now 2 modalities for facial landmark tracking: {0,1} -> {performance, quality}. Make sure to choose the preferred mode for your application. The default for face mesh fitting and expression estimation are 1, and the others are 0.
  • Head Pose output from the NvAR_Feature_LandmarkDetection feature is now in the OpenGL convention. Changed from X-back(towards the camera), Y-right, Z-down to X-right, Y-up, Z-back(towards the camera).
  • The sample apps now show the headpose in the OpenGL convention. The color coding of the axes is Red - X , Green - Y, Blue - Z
• NvCVImage_Transfer() now sets alpha to 255 or 1.0f when doing RGB -> RGBA. NvCVImage_CompositeRect() has a premultiplied alpha mode added
• Migrated to TensorRT 8.4.2.2
• Migrated to CUDA 11.6u1

v0.7.6.2 Release

Changelog (v0.7.6.2)

• 3D Body Pose Keypoint Tracking
  • Quality improvement in accuracy and stability
  • Property NvAR_Parameter_Config(NVAR_MODE) is now NvAR_Parameter_Config(Mode)
• Migrated to TensorRT 8.0.1.6
• Migrated to CUDA 11.3u1

v0.7 Beta

v0.7 Beta -NVIDIA GTC April 2021 Release

v0.6

v0.6 (Beta):

• 126 facial landmark detection and tracking: predicts and tracks the pixel locations of 126 human facial landmark points and the 3 degrees of freedom head pose in images or videos. In addition to the traditional Multi-PIE 68 point mark-ups, it detects and tracks more facial features including laugh lines, eyeballs, eyebrow contours and denser face shape landmarks, at ~800 FPS on a GeForce RTX 2060.
• GeForce RTX 30 Series support: extends the AR SDK GPU support to the latest NVIDIA GPUs.