AI video analytics with individual AI models
Individual AI models are created with SynaEYE from real video data. The integrated label manager, a separate GPU training server and controlled deployments connect the record creation, training, validation and operation to a continuous model life cycle.
Direct answer
How do individual AI models form?
Standard models cover many typical objects and situations. However, specific camera perspectives, small object sizes, special lighting conditions or industry-specific requirements can require your own training data.
Real video data becomes structured training data.
The integrated label manager prepares relevant image data for new training cycles. Misconceptions, new scenes and project-specific requirements become structured training data.
- Starting pointReal events, misunderstandings and new scenes
- LabellingAnnotate objects, states and own classes in a structured way
- TrainingTransfer datasets to separate GPU infrastructure into new model versions
- OperationsTesting, versioning, sharing, deploying and rewinding models when needed
Integrated label manager
Individual AI models: Carefully prepare image data.
The label manager integrated in SynaEYE connects real video data directly to the model workflow. Relevant recordings can be selected, annotated and compiled into controlled training data sets.
01 / 04
Select scenes
Adopt relevant events, error detections, non-recognitions and new situations from the existing visual material in a targeted manner.
02 / 04
Annotate objects
Annotate objects or visible states and uniquely assign them to the relevant image areas.
03 / 04
Create classes
Create your own object classes and project-specific states suitable for the respective application.
04 / 04
Share datasets
Build, review, expand and release training records for new training runs in a structured manner.
Separate technical rolls
Training and production video analytics remain separate.
SynaEYE deliberately separates compute-intensive model training from production video analytics.
The separate training server provides the required GPU compute. Production analytics workers continue to process live camera streams during training.
Project-specific classes
Align individual AI models with the visual process.
Standard models detect general object classes such as people or vehicles. Project-specific training extends this scope to include objects and states that are relevant for a concrete process.
This does not have to adapt the application to an existing standard model. The model can be specifically adapted to the respective visual process.
Separate training infrastructure
Train individual AI models on a controlled infrastructure.
The training server can be operated within its own or controlled infrastructure. Training data does not necessarily have to be transferred to external model or training services.
Outsource compute-intensive training specifically.
The training server is connected to SynaEYE and takes over the computationally intensive training and validation processes for the prepared data sets.
Plan performance suitable for the model project.
The required GPU performance depends on model architecture, dataset size, image resolution and the desired training duration.
GPU-based training, validation and development of new model versions.
Released models analyse the connected camera streams during operation.
Vision Language Models work on contextual and freely formulated visual tasks.
Versioning and Rollback
Review and release new model versions through a controlled process.
A new model version does not automatically replace the production version. SynaEYE keeps model versions separate and makes changes traceable.
Validation
New model versions are tested with defined positive, negative and borderline cases before production deployment.
Compare
Hit quality, false alarms, non-detections, runtime and resource requirements of different model states can be compared.
Trace
Dataset, training run, model version, validation, release and deployment remain assigned to a defined model status.
Rolling back
If a new version does not meet production requirements, the system can return to a previously approved model version.
Continue using existing data sets
Integrate YOLO and COCO into the model workflow.
Existing training data does not need to be rebuilt. Existing annotations can be adopted into the model workflow, checked and supplemented with new project-specific scenes.
Pictures and Bounding Boxes.
Existing YOLO records with footage, classes and bounding boxes can be integrated into the controlled dataset process and specifically expanded.
Complement classes and metadata in a targeted way.
Coco records with classes, annotations and metadata can be used as a starting point and supplemented with new real images from the respective operating environment.
Continuous model improvement
Feed production findings back into the model workflow.
Cameras, lighting conditions, products, workflows or environments can change. A model trained therefore does not automatically remain permanently optimal.
Record
Record detection quality, new situations, false detections and missed detections in production.
Select
Identify relevant scenes and adopt them specifically for the extension of the data set.
Train
Complement Dataset, train new model version and validate it independently against defined test data.
Rolling out
A validated model version controls the intended analysis workers and return to a previous stand if required.
Data quality and responsibility
Individual AI models: Check the quality of real images.
The decisive factors are data quality, representativeness, annotations, class demarcation and independent validation.
Representative application data
The training record must sufficiently map the real conditions of later use. If there are relevant perspectives, lighting conditions or borderline cases, the model can only process these situations to a limited extent later on.Consistent annotations
False or inconsistent annotations are transferred to the training. Classes, label rules and quality controls must therefore be clearly defined.Checking generalization
A model can deliver very good results on the existing training data and still lose quality for new cameras, environments or object variants.Validate the Real
New model versions should not enter production based solely on improved training metrics. Validation under real operating conditions is decisive.Assess data protection requirements for your specific use case. For EU locations, refer to the GDPR and the EDPB guidelines on video surveillance. Also consider the requirements of the EU AI Act, where applicable to your use case.
For private operators in Switzerland, the FDPIC guidance on video surveillance guidance.
Suitable platforms
Implement your own AI models with SynaEYE.
Training, testing and production deployment belong together. The individual tasks remain clearly assigned in the model workflow.
Own models in operation
From labelled image to production model
SynaEYE connects the Label manager, datasets and versioned models to production analytics. A separate training server provides the compute for model training.
The right choice if you:want to develop your own object classes, check with real scenes and want to roll up approved models for your analysis workers in a controlled manner.
Frequently asked questions
Labelling and model training explained briefly.
What is the SynaEYE Label manager?
The integrated label manager prepares relevant image and video data directly for the further development of AI models. Objects and states are marked, classes are defined and data sets up in a structured manner.
When does a standard model need additional training data?
Additional data makes sense if camera perspective, light, object size, environment or industry-specific situations differ from the previous training conditions.
Which own classes can be trained?
Clearly visible and consistently definable classes can be coercived, such as products, tools, machine components, packaging, protective clothing, special vehicle types or process conditions.
What is the task of the separate SynaEYE training server?
The training server connected to SynaEYE provides GPU compute for in-house training with labelled datasets. It remains separate from production analytics workers so that training does not affect live AI video analytics.
Can model training take place on its own infrastructure?
Yes. The training server can be operated within its own or controlled IT environment. Training data does not necessarily have to be transmitted to external training services.
Are hardware and GPU performance included in the training server licence?
No. Hardware and GPU computing power are not part of the SynaEYE training server licence. They are dimensioned separately according to the required training performance.
What makes training servers, SynaEYE Worker and VLM Server different?
The training server trains its own models. SynaEYE Workers productively run approved models on camera streams. A VLM server processes Vision Language Models for contextual visual analysis.
Does SynaEYE support existing YOLO and COCO data sets?
Existing data sets in common formats such as YOLO and COCO can be integrated into the model workflow, checked and supplemented by new project-specific data.
How do model versioning and rollback work?
New models are tested and approved as traceable versions. If necessary, you can return to an earlier stable model version without overwriting the previous version uncontrollably.
From video dataset to production model
Develop your individual model in a controlled manner.
We check data situation, classes, target hardware, quality criteria and integration and define a realistic model cycle.
