Occlusion Detection:
SAFR now has the ability to detect faces that are occluded. Occlusion constitutes any obstruction of the key facial features such as from a scarf, hand, glasses, hair draping over the face, etc… This capability is currently integrated to accomplish two features:
To filter out any occluded faces while learning them in the wild and thus prevent storing ambiguous face references in the SAFR person database.
For example, such a feature is used when learning and memorizing players sitting at the casino table to prevent learning them with an occlusion feature such as a wineglass in front of their face which may later create recognition inaccuracies.
To update occurrence event records with better face images without the occlusion and thus increase the value of the image stored with the event for presentation and investigation purposes.
You will find the occlusion recognition switch in the Recognition tab under SAFR Preferences as well as max tolerable occlusion level adjustment for newly learned faces.
Core Face Recognition Optimizations for NVIDIA GPUs:
These optimizations enable up to 463 recognitions per second on NVIDIA GTX 1080Ti graphics cards. This is 14x more recognition throughput in comparison to the maximum achievable on 4 Core 3.4GHz Intel Xeon Skylake-SP processor. The improvement is even more pronounced when all face attributes are computed together (identity, age, gender, sentiment). In such case optimization delivers 320 combined recognitions per second which is 40x more throughput in comparison to maximum achievable on 4 Core 3.4GHz Intel Xeon Skylake-SP processor. These optimizations also reduce recognition latency by 50% and thus enable even faster and more reliable recognition. All this results in cost reductions for on-premises core recognition subsystem deployments from $2,477 to $518 per 100 recognitions per second and from $10,667 to $797 for 100 all-attributes recognitions per second.
Note that these optimizations introduced a necessary one-time GPU calibration step which is performed when the system is started for the first time with GPU(s) present. It takes about 3 minutes per recognition model (15 minutes total) and per GPU for the system to be properly calibrated. Until this is completed, you will see System Initializing message in video view and recognition will not be be operational.
Person Body Detection NVIDIA GPU Optimizations
Person detection speed was improved by 30% and throughout by 50%. This means person detection is faster and more fluid than before. Maximum person detection throughput for our max accuracy model is 115 frames per second on NVIDIA GTX 1080Ti and 329 frames per second on NVIDIA Quadro RTX 6000. Maximum person detection throughput for our max speed model is 625 frames per second on NVIDIA GTX 1080Ti and 1052 frames per second on NVIDIA Quadro RTX 6000.
Customizable options were added to our popular traffic dashboard (available from the Reports tab in the System Console). These options enable traffic dashboard to be customized in color, logo, language, and time-range. The traffic dashboard can now also be linked directly from another web site and all customization options are available as URL query parameters. This feature enables easy integration of the dashboard into customers' portals who may wish to display the dashboard in colors and logos of their brand.
A new attendance dashboard was added to the Report tab in the System Console. For a specified time range and location, it displays all recognized individuals in attendance along with the time interval they were observed present. This dashboard can be used as a replacement of punch-card system that tracks employee attendance when properly combined with entry and exit camera monitoring ingress and egress at the work site.
Installer has been equipped with more customizable options to allow SAFR Logs to be removed from deployment and heap auto-configure behavior to automatically scale memory allocation for SAFR based on system memory available. These options enable SAFR Platform to be deployed on very small PCs (8GB RAM, 32GB Disk, $550) that can independently monitor 2 1080p video feeds. For example, such a small configuration could be used for a small SAFR Platform deployed at a casino table. The heap auto-config also enables SAFR to scale up on larger system and thus reliably handle higher recognition throughput and event traffic.
To further protect privacy, SAFR now also limits retention of system logs associated with events to the same time frame as configured for events retention in the SAFR database. This means that no trace of individual whereabouts is kept beyond the configured retention time. Recognition logs have also been reduced in their default logging level so as not to include any personally identifiable information (PII).
A small follow-up update was released later in August.