Model Training

Meridian VMS supports fine-tuning the detection model using images collected from your own cameras. By flagging false detections on the Events page and reviewing them in the training interface, you build a dataset that is specific to your environment. The system then fine-tunes the base model, producing a custom model that reduces false positives and improves accuracy for your deployment.

Training Workflow

The training process follows a four-stage cycle:

1. Flag False Detections

On the Events page, operators can mark individual detection events as false detections. Each flagged event captures the detection thumbnail image along with the bounding box data and class label.

Tip

Encourage operators to flag false detections as part of their daily workflow. The more flagged examples the system collects, the more effectively the model can be fine-tuned.

Flagging is non-destructive — the event record is preserved and simply marked with a false_detection flag. Flagged events remain visible in the event list with a visual indicator.

2. Review Flagged Images

Navigate to Settings > AI Detection to access the training review interface. This page displays all flagged detection images collected from all recording servers.

For each flagged image, the reviewer can:

• Keep the detection box — Confirm that the bounding box is correct (this was not actually a false detection, or the box should be used as a positive training example).
• Remove the detection box — Confirm that the detection was indeed false and the box should be treated as a negative example.
• Skip — Leave the image for later review.

Note

Training images are collected from all recording servers that have detection enabled. The review interface aggregates flagged images from the entire deployment into a single queue.

3. Train the Model

Once a sufficient number of flagged images have been reviewed, initiate a training run. Training can be triggered manually or configured to run on a schedule.

Scheduled training:

• Configure the training schedule in Settings > AI Detection.
• Set the preferred time of day for training to run (e.g. overnight when GPU load is lowest).
• The system automatically collects all reviewed flagged images, prepares the training dataset, and fine-tunes the base model.

What happens during training:

1. Flagged and annotated images are collected from all detection-enabled recording servers.
2. The images are prepared as a training dataset with corrected annotations.
3. The base detection model is fine-tuned using the collected dataset.
4. The resulting custom model weights are saved.

Caution

Training uses the GPU on the recording server. During training, detection throughput may be temporarily reduced. Schedule training during off-peak hours to minimise impact on live detection.

4. Distribute the Trained Model

After training completes, the custom model is distributed to all detection-enabled recording servers:

1. The trained model file is uploaded to the management server.
2. Each recording server’s detection engine is notified to download and load the new model.
3. The detection engine loads the custom model, replacing the base model for inference.

The custom model takes precedence over the base model. If the custom model file exists on a recording server, it is loaded automatically on startup.

Training Archive

The training archive is a persistent collection of all flagged and annotated images used across training runs. Key properties:

• Persists across model switches — If you change the base model (e.g. from Small to Medium), the training archive is retained. A new training run will fine-tune the new base model using the existing archive plus any new flagged images.
• Cumulative — Each training run adds to the archive. Images from previous training runs are included in future runs, providing a growing dataset that continually improves the model.
• Per-deployment — The archive is specific to your Meridian VMS deployment and reflects the unique characteristics of your camera environments.

Note

The training archive grows over time. For very large deployments with extensive flagging, monitor disk usage on the management server where the training archive is stored.

Training Effectiveness

The effectiveness of model training depends on several factors:

Factor	Impact
Number of flagged images	More examples lead to better generalisation
Diversity of environments	Images from different cameras, lighting, and weather improve robustness
Accuracy of annotations	Correctly marking true vs. false detections is critical
Training frequency	Regular training incorporates recent environmental changes

Recommended minimums:

• At least 50 flagged images before the first training run
• A mix of true positives (correctly kept) and false positives (correctly removed)
• Images from multiple cameras and times of day

Configuration Reference

Setting	Location	Description
Training schedule	Settings > AI Detection	Time of day for automatic training
Manual training trigger	Settings > AI Detection	Start a training run immediately
Review queue	Settings > AI Detection	List of flagged images awaiting review
Custom model status	Settings > AI Detection	Shows whether a custom model is active

Limitations

• Training requires a GPU on the recording server where the training job runs.
• Training does not add new object classes — it fine-tunes detection accuracy for existing classes.
• The training dataset should contain a balanced mix of positive and negative examples for best results.
• Very small datasets (fewer than 20 images) may not produce meaningful improvements.