»Imagine you could simply apply and operate AI everywhere.«
Möchten sie mehr darüber erfahren, wie AI Ihrem Unternehmen nützen kann? Dann lassen Sie uns sprechen.
ai.cowboys nerd talk episode #1
What can you get for EUR 900,563.55 in Germany?
Well, we suspect a lot. But that’s, among other things, the amount you’d spend on average for mobile data transfer if you sent a constant stream of data from a single camera at 8 frames per second (fps) at one megapixel to the cloud for 60 months. You do not believe that? Let’s take a look at the math:
8 fps * 60 sec * 60 min * 24 hrs * 365 days * 5 yrs = 1.261.440.000 frames
A megapixel image is about 300 kilobytes, which leads to the following equation:
1.261.440.000 * 300Kbps = 360.900 GByte
According to Cable UK the the average cost of mobile data transmission in Germany in Germany in 2021 will be $2.67 per Gbyte. This leads to:
360.900 Gbyte * $2,67 / 1,07 EUR = 900.563,55 EUR
An unrealistic scenario? Maybe. But:
- Even if the price were to drop by 95%, you would still have to spend EUR 750 per month and camera just for the transmission costs. Depending on scale, that very likely means the business case is out the window, if not by transmission costs, then by cloud inference costs,
- It’s very likely that you’re transferring tons of data with limited analytical value, which you won’t find out until after you’ve paid for the transfer and inference in the cloud, and
- If your company is subject to GDPR, your data privacy officer might ask you questions like: “Is there a hypothetical possibility that people in motion or any facial features are visible on the video stream?” If the answer is “Yes, …”, the hurdle for sending data to the cloud has just been raised to “almost impossible to cross”, as a person’s movement patterns and facial features are considered personally identifiable information (PII) when it comes to AI.
This list is not exhaustive, but lists three of the biggest issues when it comes to cloud-hosted computer vision for scaled edge deployments.
So what to do? A safe option would be to run the image processing AI in or near the camera and provide only textual metadata for downstream applications while discarding the original images at run-time. The typical optimization rate is over 99%. If snapshots are needed for documentation, the AI could simply send a copy with blurred PII for documentation purposes along with the metadata.
Today, there are numerous vendors of AI-enabled cameras or edge compute devices that have sufficient computing power to run complex computer vision use cases for object detection and tracking in real time. Depending on the latency requirements of the use case, even a CPU designed for edge use may be sufficient. In one of our next ai.cowboys Nerd Talk episodes, we will give benchmarks for various CPU and GPU based edge compute devices in conjunction with computer vision use cases.
Today, there are numerous vendors of AI-enabled cameras or edge compute devices that have sufficient computing power to run complex computer vision use cases for object detection and tracking in real time. Depending on the latency requirements of the use case, even a CPU designed for edge use may be sufficient. In one of our next ai.cowboys Nerd Talk episodes, we will give benchmarks for various CPU and GPU based edge compute devices in conjunction with computer vision use cases.