July 2021 Release Notes
- Enhanced upload and snap face
- Added support for *.webm and *.mkv video formats
- Also added support for the VP8 video codec
- Enhanced the person record GUI
- Improved event image quality selection and updates
- When transitioning from match < 100% to match >= 100%, updates events immediately
- Recognized images with a higher % are always better than ones with lower % match
- Fully visible faces (i.e. not clipped by edge of the screen) are considered of higher quality than clipped faces.
- Added keyboard short-cuts to search results GUI
- Cursor keys navigate the candidate grid.
- Enter key displays Face image/Scene blow-up dialog of the selected candidate.
- ESC closes the Face image/Scene blow-up dialog
- Enhanced person record changes audit tracking
- Site and source where the last edits were made are now tracked.
- Enhanced RGB Liveness accuracy
- Texture model increased its true positive rate
- Optimized small input size face detection for 16:9 video
- 16:9 aspect ratio 320x180 input size face detection model was added (in addition to 320x240)
- Enabled video viewing for Monitor role users in Operator Console
- Enhanced face detector accuracy on non-standard aspect ratios
- Aspect-fit pre-scaling is now applied this eliminating aspect ration distortions and enhancing face localization and thus recognition accuracy.
- Added support for accelerator.gpu-id feed property in VIRGO
- This property allows for explicit assignment of feeds to GPU on multi-gpu systems
- Admin API to obtain all directories in use by tenant
- Added ARES command line option (-s) for saving credentials
- Credentials are saved in SAFRActions.config in encrypted form.
- Added 3D liveness model support
- Compatible with Windows support for Intel's RealSense cameras
- Added non-FIFO queue handling to Queue Dashboard
- Enhanced server fall-back face detector accuracy.
- 480x480 retinaface detector is used as fall-back detector in HTFS
- This enhances enrollment reliability from images via web-console and APIs.
- It also enhances biometric event indexing reliability.
- It increase use of GPU memory usage by ~250GB (overall - not per feed).
- It can disabled in face service config file via face_detection_fallback=false to reduce GPU memory usage.
- Reduced MongoDB default RAM (cache) footprint:
- MongoDB cache configuration for different system RAM sizes:
- [0 .. 16GB> RAM: set cache to 1GB
- [16GB .. 48GB> RAM: set cache to 1/3 RAM - 1GB
- [48GB .. > RAM: set cache to 1/2 RAM - 1GB (MongoDB default config)
- Enhanced display of wait times in Queue Dashboard
- Added face selection query parameters
- Enhanced event server indexes/query speeds
- Enhanced access logs with client origin IP addresses
- Enhanced system auto-recovery robustness on MongoDB out of memory conditions
- Enhanced robustness of COVI start and re-start when MongoDB is offline
- Upgraded Apache for security patches
- Enhanced reliability of video file processing by VIRGO
- Starting video frames are now ensured not to be skipped due to slow initialization.
- Added RBG Liveness support
- Added RBG Liveness Action support
- Added Secure Access with Smile and RBG Liveness support
- Added support for *.webm and *.mkv video formats
- Also added support for the VP8 video codec
- Reduced MongoDB binary size:
- All of the Windows Platform changes with the following exceptions:
- Windows Desktop Client changes aren't included.
- 3D liveness model support isn't included.
- All of the Linux Platform changes
- Android SAFR 3.5 enhancements
- Enabled imageTime event attribute
- Updated the default mask threshold to 0.5. MaskCheck remains at 0.7.
- Added a new global URLSessionDelegate that can be used to add IPAddress certificate exceptions for all the sessions in the app.
- If the similarity score goes over 100% and the current similarity score for the event is less than 100% then update the image. This can only occur once. This was ported from Windows.
- The age, gender, occlusion, and mask icons are not supposed to show up in "Detect Mask" mode. In this mode SAFR is meant to work very similar to MaskCheck. When in "Detect Mask" mode the detect mask action is also enabled.
- The source rotation of videos is now handled. The metadata is now checked to get the display matrix and then get the rotation angle. The source rotation is now saved off.
- Enabled the imageTime event attribute
- Added aspect-fit pre-scaling step for high sensitivity face detector
- Increased detection accuracy on non-standard image sizes (aspect-ratios).
- Added support for 640x360 and 360x640 aspect ratios.
- Increased accuracy of the high sensitivity quantized models.