Skip to main content

Object Classification

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.

info

Training a custom object classification model requires an internet connection to download MobileNetV2 base weights. By default these weights are not cached in /config/, so they are downloaded again after the container is recreated. Once trained, the model runs fully offline. See Network Requirements for details.

Minimum System Requirements​

Object classification models are lightweight and run very fast on CPU.

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.

A CPU with AVX + AVX2 instructions is required for training and inference.

Classes​

Classes are the categories your model will learn to distinguish between. Each class represents a distinct visual category that the model will predict.

For object classification:

  • Define classes that represent different types or attributes of the detected object
  • Examples: For person objects, classes might be delivery_person, resident, stranger
  • Include a none class for objects that don't fit any specific category
  • Keep classes visually distinct to improve accuracy

Classification Type​

  • Sub label:

    • Applied to the object's sub_label field.
    • Ideal for a single, more specific identity or type.
    • Example: cat → Leo, Charlie, None.

  • Attribute:

    • 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>.
    • Ideal when multiple attributes can coexist independently.
    • Example: Detecting if a person in a construction yard is wearing a helmet or not, and if they are wearing a yellow vest or not.
note

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.

Assignment Requirements​

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.

Example use cases​

Sub label​

  • Known pet vs unknown: For dog objects, set sub label to your pet's name (e.g., buddy) or none for others.
  • Mail truck vs normal car: For car, classify as mail_truck vs car to filter important arrivals.
  • Delivery vs non-delivery person: For person, classify delivery vs visitor based on uniform/props.

Attributes​

  • Backpack: For person, add attribute backpack: yes/no.
  • Helmet: For person (worksite), add helmet: yes/no.
  • Leash: For dog, add leash: yes/no (useful for park or yard rules).
  • Ladder rack: For truck, add ladder_rack: yes/no to flag service vehicles.

Configuration​

Object classification is configured as a custom classification model. Each model has its own name and settings. Specify which object labels should be classified.

Navigate to the Classification page from the main navigation sidebar, then click Add Classification.

In the Create New Classification dialog:

FieldDescription
NameA name for your classification model (e.g., dog)
TypeSelect Object for object classification
Object LabelThe object label to classify (e.g., dog, person, car)
Classification TypeWhether to assign results as a Sub Label or Attribute
ClassesThe class names the model will learn to distinguish between

The threshold (default: 0.8) can be adjusted in the YAML configuration.

Training the model​

Creating and training the model is done within the Frigate UI using the Classification page. The process consists of two steps:

Step 1: Name and Define​

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.

Step 2: Assign Training Examples​

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.

Improving the Model​

Diversity matters far more than volume

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.

  • 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.
  • Problem framing: Keep classes visually distinct and relevant to the chosen object types.
  • Preprocessing: Ensure examples reflect object crops similar to Frigate's boxes; keep the subject centered.
  • 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).
  • Threshold: Tune threshold per model to reduce false assignments. Start at 0.8 and adjust based on validation.
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.

Debugging Classification Models​

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.

Navigate to Settings→System→Logging.

  • Set Logging level to debug
  • Set Per-process log level > frigate.data_processing.real_time.custom_classification to debug for verbose classification logging

The debug logs will show:

  • Classification probabilities for each attempt
  • Whether scores meet the threshold requirement
  • Consensus calculations and when assignments are made
  • Object classification history and weighted scores