Make and Model Recognition (MMR)

Purpose-built, high-accuracy AI for detailed road-user and vehicle recognition — deployable from embedded edge devices to servers and cloud.

Highlights

15

categories

374

makes

3085

models

8030

generations

23

TAGs

day/night

light

360°

views

Accuracy & Performance

99%

category

99%

make

98%

model

95%

generation & variation

4 ms

GPU speed

Developed since

2007

What MMR recognizes

Eyedea MMR is an advanced AI-based vehicle recognition and classification system that identifies detailed road-user and vehicle attributes directly from an image. Recognition is based on the visual appearance of the road user itself and does not depend on license-plate information.

  • Road-user type
  • Vehicle category
  • Make
  • Model
  • Generation
  • Variant
  • Color
  • Tags & visual attributes
  • Axles, wheels, roof cargo & other details
  • Vehicle view

Accuracy & coverage

Recognition accuracy
Feature Number of Classes Accuracy
Road User Type & Vehicle Category 15 > 99%
Make 374 > 99%
Model 3,085 > 98%
Generation 8,030 > 95%
Variant 3,470 > 95%
Color 12 > 80%
Tags 23 > 98%
Accuracy depends on image quality, vehicle visibility, region and other acquisition conditions. See the MMR Technical Sheet for detailed benchmark methodology.
 
Operational coverage
Camera Views Frontal, Front-Side, Side, Rear, Rear-Side, Top, Bottom
Camera Types Static (Fixed), PTZ (Pan-Tilt-Zoom), Mobile (Moving), Handheld
Regions Europe, Middle East, North & South America, Asia-Pacific, South Africa
Lighting Day and Night
Platforms
  • Ubuntu (x86_64, CPU | GPU)
  • Ubuntu (AArch64, CPU | GPU)
  • Windows (x64, CPU | GPU)
  • NVIDIA Jetson (GPU)
  • HAILO
* Other platforms on request
Deployment
  • Embedded Edge Devices
  • Servers
  • Cloud
Performance ~4 ms per vehicle on GPU, ~8 ms on CPU
See the MMR Technical Sheet for detailed benchmarking.

Deploy MMR your way

MMR SDK Package

MMR SDK is the foundation of our advanced Make & Model Recognition technology. Designed for maximum flexibility, it integrates seamlessly into products ranging from lightweight embedded edge devices to high-performance servers.

With ready-to-use libraries, cross-platform support, and clear documentation, MMR SDK empowers businesses to accelerate development and bring powerful recognition features to market faster.

The SDK includes:
- A shared library (.so, .dll) with a C interface
- Wrappers for Java and Python
- A sample project to get you started quickly
- Comprehensive documentation for smooth implementation

Download the latest MMR SDK packages from this SharePoint folder or use the provided links.

Latest packages: v2.27.x (released 2026-06-25)

Typical applications

  • e-tolling
  • traffic safety and surveillance
  • law enforcement
  • police and governmental use
  • access control and parking management
  • traffic infrastructure planning
  • smart city solutions
  • fleet & insurance applications

How it works

The core of MMR software is a deep neural network, specifically designed and trained by Eyedea for the task of detailed recognition of road users and vehicles. The network takes a selected object crop as input and outputs detailed characteristics of the analyzed object. The evaluation of input data by the neural network (inference) is implemented within the MMR SDK, which is subsequently utilized across our other products, including Desktop and Console application Traffic Analyzer, MMR+ANPR REST Server, and Cloud service TrafficEye.AI.

We recommend using MMR technology in combination with Eyedea object detectors, which are included in the Eyedea ANPR packages. However, it can also be integrated with custom or third-party object detectors.

Eyedea MMR SDK architecture showing vehicle image input, neural network inference, structured vehicle attributes and deployment options.
Eyedea MMR processing pipeline and deployment options.

Latest MMR developments

Vehicle Identification from Images

Eyedea MMR can identify vehicles directly from photographs and video frames. Try it online and see detailed vehicle recognition results together with license-plate information.

Try Vehicle Recognition & Number Plate Reading Online

Upload an image and see vehicle recognition and number-plate recognition results instantly. Your first 20 images are free.

Frequently Asked Questions:

What does MMR stand for in the automotive context?

MMR stands for Make and Model Recognition. It is an AI-based vehicle recognition technology that identifies a vehicle’s category, make, model, generation, variant, and color from an image. In Eyedea’s solution, MMR can also detect selected vehicle tags, such as taxi, ambulance, pickup, or fire brigade vehicle.

How does car identification from a photo work?

Car make and model identification is performed using AI image analysis. For car model identification from photo, a cropped image of the target road user is processed by a deep neural network specifically designed and trained for this task by the Eyedea team. The network evaluates information from all pixels in the image in a process known as inference. As a result, the system can determine the car model from image and return all recognized characteristics together with their confidence scores, which express the probability that each result is correct.

Does MMR technology work with still images, videos, or both?

MMR technology can work with both still images and video, depending on how it is integrated. At its core, the MMR-SDK processes a cropped image of the target road user and uses a deep neural network to analyze its visual features. From a single photo, the system can identify the vehicle’s category, make, model, generation, variant, color, and selected tags, and return these results together with confidence scores. Since a video is simply a sequence of still images, the same recognition can be performed frame by frame on a video stream. This approach is implemented in our Traffic Analyzer application, which is designed for video analysis. By contrast, the MMR+ANPR REST Server application and the TrafficEye.AI cloud service are designed to process still images only.

Does it run locally on embedded edge devices, or is it cloud-based?

MMR can be used on a wide range of platforms, from embedded edge devices (e.g., directly in traffic cameras) to large back-office servers and cloud environments. The software is designed to run on CPUs (ARM, x86_64, etc.), on GPUs (both Nvidia and non-Nvidia), and on AI accelerators (HAILO, DLA, and others).

Where can I find a list of recognized vehicle makes and models?

MMR recognizes hundreds of makes and thousands of models, generations, and variants, so we cannot list them all here. The complete list of recognized makes, models, generations, and variants is provided in the CSV file: MMRBOXVCMMGVCTPREC_2025Q2.MMR.csv.
This file is included in the MMR SDK package in the following folder: \sdk\modules\[edftf2lite, edftrt]\model\

How does LP-based MMR differ from Road-User Box-based MMR?

ANPR and MMR serve different purposes. ANPR is used to detect and read the license plate, while MMR is used to analyze the vehicle itself and recognize attributes such as category, make, model, generation, variant, and color.

In earlier MMR deployments, the position of the target vehicle was derived from the license plate location provided by the ANPR system. Because the license plate position was already available, there was no need for a separate vehicle detector, which reduced computational cost. This approach is called LP-based MMR.

In current Eyedea systems, the detector can localize the entire road user bounding box together with sub-parts such as license plates, windshields, and axles in a single step. That means MMR can use the full object bounding box directly as the input reference, rather than relying only on the license plate position. This approach is called Road-User Box-based MMR.

Both approaches are still supported for compatibility with different integration scenarios.

Does MMR also work with night images?

Yes, MMR is trained to analyze night images as well, without any significant drop in accuracy or performance. For more details, please refer to the benchmark tests in the MMR SDK Technical Sheet.

Does it work with color images only, or can monochrome images also be processed?

MMR can process both color and grayscale images with the same level of accuracy and performance. The only limitation is that vehicle color cannot be recognized from a grayscale image.

Does the TrafficEye.AI cloud service read license plates that are illegible to the human eye?

No. For such cases, we offer a separate product called Forensic ANPR, which has been specifically developed to read illegible plates.

Is the MMR-SDK capable of detecting (localizing) vehicles on its own?

No. The MMR-SDK does not include an object detection or localization module. Vehicle detection is provided as part of the ANPR SDK package. This function returns bounding boxes for all types of road users — including pedestrians, cyclists, motorbikes, and vehicles — as well as for specific vehicle components such as windshields, axles, and number plates. We recommend using this function to determine vehicle bounding boxes. However, thanks to the robustness of the MMR module, it is also possible to use a third-party object detection solution if preferred.

How long has Eyedea been involved in MMR development?

Eyedea is a pioneer in this field. The first version of the MMR product was launched in 2007 under the name Vehicle Classification (VCL). At that time, the solution relied on completely different algorithms: object detection was implemented using the WaldBoost algorithm, while vehicle classification was based on multi-class Support Vector Machines (SVMs). Only the vehicle category, make, and color were recognized, and the system required daylight conditions.

With the introduction of deep neural networks in 2015, Eyedea significantly expanded the capabilities of MMR. The number of supported classes grew from tens to thousands, night-time and multi-view recognition were added, and the system evolved to provide detailed classification including model, generation, variant, tags, and other vehicle attributes.

MMR continues to evolve through two major updates each year, adding newly launched vehicle models, expanding regional coverage, introducing new recognized attributes, and further improving recognition accuracy and runtime performance.

Looking for a vehicle recognition solution?

Our references