Site-specific objects
Dedicated tools, equipment signs, industrial materials and other objects with a recognisable appearance.
Custom computer vision models
XINHUOAI Model Training Platform
Prepare field samples, train a model, inspect the errors and deploy a compatible version to your local AI device. XINHUOAI supports the model-production work; day-to-day recognition runs on the selected deployment hardware.
When a project needs a new model
Use a standard algorithm where it already meets the requirement. Custom training is useful when the camera can clearly see a relevant target, but the existing model does not recognise it reliably or does not cover that target class.
Workwear and PPE variants
Uniform colours, garment designs and visible protective equipment that differ from the existing model's training examples.
Visible equipment states
Observable changes in an indicator, component or work area. Confirm that the required condition is visible at the installed camera distance.
Defined scene events
Object presence or absence, placement conditions and clearly defined actions. Temporal events may need a sequence model and rules as well as image detection.
Training cannot recover a detail that the camera does not capture. Vague intent, hidden actions or precision measurements beyond the image resolution need a different sensing method or a narrower requirement.
Prepare data that matches the deployment
Send footage from the intended cameras and describe what should trigger an event. Useful coverage includes normal activity as well as the rare situations the project is meant to detect.
| Material | Include | Purpose |
|---|---|---|
| Positive examples | Visible targets at different distances, angles, lighting conditions and partial occlusions. | Represent how the target actually appears on site. |
| Negative examples | Look-alike objects, reflections, shadows, empty scenes and normal operating activity. | Test and reduce false detections. |
| Video clips | Target arrival, movement, temporary occlusion and departure, including relevant time periods. | Check continuity and event rules on real footage. |
| Scene notes | Camera height and angle, target size, detection region, schedule, event definition and intended hardware. | Connect the model task to the operating requirement. |
| Field errors | False alarms and missed events with their source camera and circumstances. | Guide the next dataset and model iteration. |
Use material you are authorised to supply. Agree data access, transfer, retention and any redaction requirements before uploading project footage.
Training and quality checks
Define the task
Record target classes, expected output, camera conditions, deployment device and evaluation criteria.
Clean the dataset
Remove unusable images and redundant duplicates. Keep the source, scene and collection context of useful samples.
Review annotations
Check class names, target boxes or other labels. Existing models can assist with pre-annotation where suitable; people still review the labels.
Separate the datasets
Keep training, validation and final test material distinct. Avoid near-identical video frames appearing on both sides of a test.
Train and evaluate
Use GPU resources for training, then inspect precision, recall and task-appropriate metrics alongside the actual errors.
Validate on hardware
Export or convert the model, check inference performance and test the end-to-end event workflow on the intended device.
Review the errors, not just the overall score
A useful evaluation explains which scenes work, which fail and under what conditions. Raising a confidence threshold may reject false detections but can also miss real targets. Review both sides before accepting a change.
Model-level evaluation
Inspect class-level precision and recall, false positives, missed targets and the sample coverage behind the result. Use metrics that fit the task, such as F1 or detection mAP where appropriate.
Event-level verification
Test the video source, model, region rules, dwell conditions, evidence capture and platform delivery together. Record nuisance alarms, missed events and response time over representative operating periods.
Keep difficult field examples for the next iteration. Repeat validation when camera position, lighting, visible targets or operating rules change.
Deploy to hardware that fits the model
| Target | Typical role | Compatibility checks |
|---|---|---|
| AI camera | Recognition at an individual camera point. | Model format, supported operators, device memory and inference time. |
| AI Edge Box | Local analysis of several accessible camera streams. | Model conversion, resolution, channel count and concurrent algorithm load. |
| AI NVR | Recording with supported AI events and evidence search. | Supported software version and integration with the NVR event workflow. |
| Video analytics server | Larger video workloads or project-specific model services. | GPU or accelerator support, memory, decoding capacity and service workload. |
A model export is not automatically compatible with every device. Conversion, quantisation and runtime validation may be required. Once a compatible model is deployed, local recognition does not require the training platform to remain connected.
Training resources and supporting services
XINHUOAI supports customer-led material upload, training tasks, result review and model export. Projects that need assistance can include data annotation, model tuning and deployment verification.
- GPU training time
- Training resource usage and duration form the basis of the training charge.
- Data annotation
- Sample cleaning, target labels and annotation review can be scoped as a service.
- Model tuning
- Additional work addresses the agreed field errors and difficult scene conditions.
- Deployment support
- Conversion, device testing and on-site or remote integration are scoped for the target hardware.
The quotation should identify the target classes, dataset work, GPU resources, deployment devices, deliverables and acceptance method. No fixed training fee or guaranteed accuracy applies to every possible scene.
Model training questions
Can we supply our own project images and videos?
Yes. Provide authorised samples from the intended cameras, including normal scenes, target examples and confusing negatives. Include an explanation of the event you need to detect.
Can we train without an in-house algorithm team?
The platform supports a customer-led workflow. Data quality, task definition and evaluation still need attention; XINHUO can scope annotation, tuning and deployment support for the project.
How many images do we need?
There is no single number for every task. Required coverage depends on target variation, classes, camera views, lighting and the difficulty of negative examples. Review representative samples before deciding the collection plan.
Will the model work on any AI camera or Edge Box?
No. The export must match the selected runtime and hardware. Model format, operators, memory use, resolution and inference performance need validation before deployment.
Does offline recognition mean training is also offline?
No. Supported models can run locally after deployment, while the training service has its own resource and connectivity requirements. A private training deployment is a separate requirement to confirm.
Can we update the model after installation?
Yes, through a compatible model update and validation process. Collect field errors, prepare a new dataset iteration and test the updated model before rolling it out across sites.
Project inquiry
Tell us your camera environment and required AI analytics.
Tell us the camera count, the events to detect and the project location. Our team will reply by email with the relevant product information and next steps.