Calculating video memory for analytics
Our face and license plate recognition modules are based on neural networks. We have created a separate neural network for each type of recognition consuming custom amount of video memory (VRAM). The analytics module video memory consumption depends on the type of recognition you want to use. You can use the following figures in your calculations:
- License plate recognition consumes 1.7 GB of video memory per server and about 150 MB per each camera
- License plate recognition with detection of vehicles without license plates consumes 4.4 GB of video memory per server and about 150 MB per each camera
- Emergency vehicle detection consumes 4.4 GB of video memory per server and about 150 MB per each camera
- Face recognition consumes 2.5 GB of video memory per server and about 200 MB per each camera
It is also recommended to make a memory margin of about 1 GB for each neural network.
VRAM for one type of recognition = Video memory per neural network + Number of cameras × Video memory per camera + 1 GB
Please refer to this page for other system requirements for analytics.
Let's consider the situation when Watcher controls access to some facility. For that, there are 5 checkpoints each having one camera recognizing license plates to open the barriers, and 10 checkpoints each having one camera for face recognition to open the doors. Assume that we connect all these cameras to one streamer which will perform both license plate recognition and face recognition.
VRAM required for license plate recognition: 1,7 GB + 5 × 150 MM + 1 GB ≈ 3,4 GB
VRAM required for face recognition: 2,5 GB + 10 × 200 MB + 1 GB ≈ 5,4 GB
Total VRAM: 3,4 GB + 5,4 GB = 8,8 GB
Knowing this number and versions of Compute Capability supported by Watcher (6.1, 7.5 and 8.6) you can choose your video card(s) on the Nvidia website