Skip to main content

Model Configuration

The Models Configuration dialog is where you set up the models used in your project. You can create multiple models, decide what each one should do, and configure how it behaves during inference and training.

Use this page as a short walkthrough. If you want a field-by-field reference, see Model Configuration Reference.

Models configuration dialog

Before you start​

Open the dialog from your project and select the model you want to edit, or create a new one.

The layout has two parts:

  • On the left: the list of models in the project
  • On the right: tabs for the selected model

The available tabs are:

  • Configuration
  • Inference
  • Categories for Locate and Classify models
  • Key points when key points are enabled
  • Attributes
  • Training

1. Add or choose a model​

Use the model list on the left to:

  • add a new model
  • switch between existing models
  • remove a model you no longer need
  • choose the project’s Primary model

The primary model is the main entry point used by the app when it starts an analysis.

2. Configure the model type​

Start in the Configuration tab.

General settings​

Fill in:

  • Name: a clear name for the model
  • Primary: marks this model as the main model for the project
  • Input width and Input height: image size used by the model
  • Description: a short explanation of what the model is for

Base settings​

Choose what the model should do:

  • Locate: finds objects in images
  • Classify: classifies images or detected objects, and can also use boxes or key points
  • Identify: identifies already-located objects

Then choose the options that fit your use case:

  • Accuracy: balances speed and quality
  • Bounding boxes: enables object locations. Models can also be set up to detect rotated objects, so tilted items get tight, angled boxes — see Box head mode in the Model Configuration Reference.
  • Key points: enables landmark detection
  • Multiple categories: allows more than one class
  • Based on: starts from scratch or from the COCO 2017 preset

When using COCO 2017, training speed is improved considerably, no matter what type of image data.

3. Set inference behavior​

Use the Inference tab to control how strict the model is when it runs.

Accuracy thresholds​

These values do not act as a single fixed threshold on their own.

They define the model's accuracy range:

  • Base Accuracy
  • Lower Limit
  • Upper Limit

That range is used together with the global find-accuracy slider in the Engine tab of the menu.

  • Base Accuracy is the model's default point
  • Lower Limit is how far the global slider can make the model less strict
  • Upper Limit is how far the global slider can make the model more strict

In practice, you usually set these values as the allowed range for the model, then use the global find-accuracy slider to move within that range while searching or analyzing.

Limits​

If the model supports multiple categories, you can set Maximum Detections to limit how many results are returned.

NMS and overlap control​

If the model uses bounding boxes, you can also configure:

  • Use NMS: turns duplicate cleanup on or off. When enabled, overlapping detections are compared and weaker duplicates are removed.
  • Pre NMS Top K: limits how many candidate detections are considered before NMS runs. Higher values keep more possible detections, especially in busy images.
  • NMS IoU Threshold: controls how much two boxes may overlap before one suppresses the other. Lower values remove overlaps more aggressively.
  • Class-Agnostic NMS: lets boxes suppress each other across different categories instead of only within the same category.
  • NMS Overlap Groups: lets you define groups of categories that should suppress each other when class-agnostic NMS is turned off.
info

NMS stands for Non-Maximum Suppression. It is a common way to clean up duplicate detections. When the model returns several overlapping boxes for what is really the same object, NMS keeps the strongest result and removes the weaker overlaps.

Use these settings when you want better control over overlapping detections. In many projects, the default settings are good enough until you start fine-tuning model behavior.

4. Define categories​

Use the Categories tab to define what the model can detect.

For each category, you can:

  • set the numeric value and display name
  • turn ROI display on or off
  • optionally override accuracy thresholds for that category
  • optionally assign the category to one or more NMS overlap groups
  • optionally set a per-category class weight when class weighting is enabled for training

This tab is where you keep the class list clean and practical. Only include categories you actually need.

warning

Changing the name of a category here can have unexpected effects for detecting events in the data.

5. Define key points​

If Key points is enabled, use the Key points tab.

This tab lets you:

  • choose which categories produce key points
  • define the ordered list of key point labels
  • define meshes that connect key points for display or structure

The order of the key point labels matters. Keep it stable once you start labeling training data.

6. Add attributes​

Use the Attributes tab to calculate extra information for detections.

There are two levels:

  • Shared attributes: apply across the whole model
  • Per-category attributes: apply only to specific categories

Common attribute types include:

  • Color
  • Another model
  • Identification
  • Border width calculator
  • Angle
  • Bearing
  • Adjacent angle

You can also choose to clip to a bounding box or align using key points when the source model supports it.

Each attribute has a Skip frames setting that controls how often it is calculated. Set it to 0 to calculate every frame, or to a higher number to calculate less frequently. This is useful for costly attributes that do not need to update every frame.

When the model has Multiple categories enabled and at least one category defined, you can also set Requires categories on any attribute. Select one or more categories and the attribute will only run when at least one of those categories is detected in the same image or frame. Leave the selection empty to keep the attribute running unconditionally.

Use attributes when you want to build a workflow such as:

  • detect a person, then identify them with another model
  • detect a face, then estimate angle or bearing
  • detect an object, then extract color information

7. Set training options​

Use the Training tab when you want to train or retrain the model.

Data preparation​

You can control:

  • Prepare data
  • Percentage training data
  • Level categories

Use Prepare data when you want the app to create a fresh training and validation split.

Clipping and alignment​

You can also set training-time:

  • clipping
  • alignment with key points
  • crop margin

These are useful when the model should focus on the object itself instead of the full image.

Data augmentation​

You can enable augmentations such as:

  • rotate
  • scale
  • horizontal or vertical flip
  • translate X or Y
  • brightness, contrast, saturation, hue
  • number of augmentation combinations

Use augmentation when you need more variation in the training data.

Core training settings​

The training tab also includes the main controls for:

  • batch size
  • epochs
  • max epochs
  • optimizer
  • momentum
  • patience
  • fine tuning
  • weight decay
  • clip norm

Learning rate settings​

Learning-rate controls are now grouped separately and include:

  • learning rate decay type
  • initial learning rate
  • decay steps
  • differential learning rates
  • head and backbone learning rates
  • warmup
  • cosine restart options

These settings are mainly for advanced tuning. Most users should start with the defaults.

Advanced model-specific settings​

When the model uses boxes or key points, extra settings become available for:

  • backbone freezing and staged unfreezing
  • batch norm behavior
  • loss weights
  • heatmap and Gaussian settings
  • CenterNet class weighting

If you are not actively tuning training behavior, it is usually best to leave these advanced values alone at first.

Genetic optimization​

You can also enable genetic optimization to search over selected training fields automatically.

note

This feature is still experimental

Saving changes​

Select Save to keep your changes.

If there are validation problems, the app will show them and stop the save until they are fixed.

If you close the dialog with unsaved changes, the app asks for confirmation.

Suggested workflow​

For a new model, a good order is:

  1. Create the model and choose its purpose in Configuration
  2. Add categories in Categories
  3. Add key points if needed
  4. Add attributes only if you need extra outputs
  5. Check Inference settings
  6. Start with default Training settings, then adjust only after a first training run