EPISODE 7. EXTERNAL PROCESSING AND HEAVY NEURAL NETWORKS

Today, the real-world solution most often looks like this: cameras capture the image, and all processing is done on an external server with a GPU. Such a server is capable of running large models that analyze not only static frames, but also temporal sequences, movement dynamics, and human interaction with water.

Here we encounter another peculiarity: “water” tasks require much more complex neural networks than, say, recognizing license plates or identifying faces. A car always looks roughly the same, and a face is a static object. In water, everything is dynamic and unstable, so the network must be “heavier” and more powerful.

Of course, there are intermediate options for compact solutions. For example, the NVIDIA Jetson family. These small modules allow you to run neural networks directly on the device, and they are much more flexible than “firmware-based” smart cameras. But there is a compromise here too: power is limited, which means you have to use more compact models.