Object classification allows you to train a custom MobileNetV2 classification model to run on tracked objects (persons, cars, animals, etc.) to identify a finer category or attribute for that object. Classification results are visible in the Tracked Object Details pane in Explore, through the `frigate/tracked_object_details` MQTT topic, in Home Assistant sensors via the official Frigate integration, or through the event endpoints in the HTTP API.
Training a custom object classification model requires a one-time internet connection to download MobileNetV2 base weights. Once trained, the model runs fully offline. See [Network Requirements](/frigate/network_requirements#one-time-model-downloads) for details.
Training the model does briefly use a high amount of system resources for about 1-3 minutes per training run. On lower-power devices, training may take longer.
- Added as metadata to the object, visible in the Tracked Object Details pane in Explore, `frigate/events` MQTT messages, and the HTTP API response as `<model_name>: <predicted_value>`.
A tracked object can only have a single sub label. If you are using Triggers or Face Recognition and you configure an object classification model for `person` using the sub label type, your sub label may not be assigned correctly as it depends on which enrichment completes its analysis first. This could also occur with `car` objects that are assigned a sub label for a delivery carrier. Consider using the `attribute` type instead.
Sub labels and attributes are only assigned when both conditions are met:
1.**Threshold**: Each classification attempt must have a confidence score that meets or exceeds the configured `threshold` (default: `0.8`).
2.**Class Consensus**: After at least 3 classification attempts, 60% of attempts must agree on the same class label. If the consensus class is `none`, no assignment is made.
This two-step verification prevents false positives by requiring consistent predictions across multiple frames before assigning a sub label or attribute.
Object classification is configured as a custom classification model. Each model has its own name and settings. Specify which object labels should be classified.
<ConfigTabs>
<TabItem value="ui">
Navigate to the **Classification** page from the main navigation sidebar, then click **Add Classification**.
An optional config, `save_attempts`, can be set as a key under the model name. This defines the number of classification attempts to save in the Recent Classifications tab. For object classification models, the default is 200.
Enter a name for your model, select the object label to classify (e.g., `person`, `dog`, `car`), choose the classification type (sub label or attribute), and define your classes. Frigate will automatically include a `none` class for objects that don't fit any specific category.
For example: To classify your two cats, create a model named "Our Cats" and create two classes, "Charlie" and "Leo". A third class, "none", will be created automatically for other neighborhood cats that are not your own.
The system will automatically generate example images from detected objects matching your selected label. You'll be guided through each class one at a time to select which images represent that class. Any images not assigned to a specific class will automatically be assigned to `none` when you complete the last class. Once all images are processed, training will begin automatically.
When choosing which objects to classify, start with a small number of visually distinct classes and ensure your training samples match camera viewpoints and distances typical for those objects.
If examples for some of your classes do not appear in the grid, you can continue configuring the model without them. New images will begin to appear in the Recent Classifications view. When your missing classes are seen, classify them from this view and retrain your model.
Selecting dozens of nearly identical images is one of the fastest ways to degrade model performance. MobileNetV2 can overfit quickly when trained on homogeneous data — the model learns what _that exact moment_ looked like rather than what actually defines the class. **This is why Frigate does not implement bulk training in the UI.**
For more detail, see [Frigate Tip: Best Practices for Training Face and Custom Classification Models](https://github.com/blakeblackshear/frigate/discussions/21374).
:::
- **Start small and iterate**: Begin with a small, representative set of images per class. Models often begin working well with surprisingly few examples and improve naturally over time.
- **Favor hard examples**: When images appear in the Recent Classifications tab, prioritize images scoring below 90-100% or those captured under new lighting, weather, or distance conditions.
- **Avoid bulk training similar images**: Training large batches of images that already score 100% (or close) adds little new information and increases the risk of overfitting.
- **The wizard is just the starting point**: You don't need to find and label every class upfront. Missing classes will naturally appear in Recent Classifications, and those images tend to be more valuable because they represent new conditions and edge cases.
- **Crop size**: Aim for crops of at least 100×100 pixels (a 10,000 pixel area). Crops smaller than ~80×80 get stretched 3-7× by the model's 224×224 input resize and tend to collapse into a generic "blob" region of feature space where identity becomes unreliable. If most of your detections are small because the camera is far from the subject, consider repositioning the camera for closer crops.
- **Class balance**: Aim to keep your largest class within ~3× the count of your smallest. Beyond that, the model becomes biased toward the dominant class and tends to default borderline predictions to it (the "everything looks like Buddy" failure mode).
:::tip `none` works differently from named classes
Named classes work best with visually uniform examples — every Buddy photo should look like Buddy. The `none` class needs the opposite: visual diversity across sizes, framings, and qualities, because at inference it has to absorb everything that isn't one of your named classes. Don't apply the same "only keep large, well-framed images" rule to `none` that you would to a named class. Mix in small crops, partial views, and false positives deliberately - otherwise the model has no signal for "small/ambiguous thing = not one of my known classes" and will force those crops into a named class by default.
To troubleshoot issues with object classification models, enable debug logging to see detailed information about classification attempts, scores, and consensus calculations.
Enable debug logs for classification models by adding `frigate.data_processing.real_time.custom_classification: debug` to your `logger` configuration. These logs are verbose, so only keep this enabled when necessary. Restart Frigate after this change.
