CarNET Server
On-premise car recognition: the CarNET recognition API on your own hardware. It works without internet access, and your photos stay inside your network.
Who It's For
- Closed networks: security, police, government, ports and factories.
- Parking and access operators with their own cameras.
- Companies whose data rules don't allow sending photos out.
- High volumes, where a price per request adds up.
What It Recognizes
The same as our Cloud API: the make, model and generation of more than 3,100 car models, plus colour and angle. It works with the worldwide and the North American car lists, and with several cars in one photo.
Requests and answers are the same as the Cloud API's, so code written for one works with the other. How recognition works explains the steps and options.
How It's Delivered
We send you CarNET Server as a Docker image in a file, so your server needs no access to the internet or to us. You load it with one command and start it, and it answers requests on your own network.
Updates with new car models arrive as files too, and you load them the same way.
Requirements
| CPU version | GPU version | |
|---|---|---|
| Server | Linux on a 64-bit Intel or AMD processor, physical or virtual | The same, plus an NVIDIA GPU from 2016 or newer |
| Software | Docker | Docker, the NVIDIA driver and the NVIDIA Container Toolkit |
| Memory | 4 GB RAM | 4 GB RAM |
| Disk | About 1 GB | About 4 GB |
Kubernetes works with the same image. We can also build CarNET Server for Windows.
CPU or GPU, and Speed
A CPU is enough for most sites. A GPU helps with many camera streams or very high volumes.
| Hardware | Time for one photo | Photos per second, fully loaded | Photos per day, fully loaded |
|---|---|---|---|
| 1 CPU core (Intel Core i5-8400, 2018) | 0.12 s | 8 | 0.7 million |
| 6-core CPU (Intel Core i5-8400, 2018) | 0.05 s | 32 | 2.7 million |
| The same, with an NVIDIA GTX 1080 Ti GPU (2017) | 0.03 s | 58 | 5 million |
Indicative figures from our own tests: one car per photo, typical photo sizes, make, model and generation. Time for one photo: one photo at a time, using the whole machine. Fully loaded: several photos at once, one per CPU core (four at a time with the GPU). Asking for colour and angle too adds 10–35%. Very large photos take longer, and newer hardware is faster.
Photos That Work Best
Best results
- A camera at about the height of a car, or a little above it.
- The car seen from the front, the rear, the side or at an angle.
- The car large in the frame: at least 180 pixels high and wide.
- The whole car in the frame, not cut off at the edge.
- Daylight or good lighting.
May reduce recognition quality
- Cameras mounted high and looking steeply down at the cars.
- Dark or infrared night images.
- Strong blur.
- Cars that are mostly hidden behind other cars or objects.
- Close-ups of a wheel, a badge or the interior: we report these as such, not as a model.
- Photos of only the number plate.
Not sure about your cameras? Contact us for a free check: you send 50–100 photos from your cameras, and we run them through our recognition and send you the results before you decide.
Updates and Support
Updates bring new car models and better recognition as they reach our Cloud API. You choose how often: monthly, quarterly, or on demand when you need certain models as soon as possible. Support is by email.
Licence and Price
Price on request. It depends on your volume, the number of servers and how often you want updates.
Your Photos Stay With You
CarNET Server runs entirely inside your network and doesn't connect to the internet. Your photos stay on your servers; we never see them.
