Per-Class Confidence Filter¶

Filter detection predictions by applying a different confidence threshold to each class, keeping only detections whose confidence meets or exceeds the threshold configured for their class (with a configurable fallback threshold for classes that are not listed).

How This Block Works¶

This block applies class-aware confidence filtering to detection predictions, enabling precise control over which detections are retained based on per-class quality requirements. The block:

1. Takes detection predictions (object detection, instance segmentation, or keypoint detection) and a dictionary mapping class names to confidence thresholds
2. Iterates through each detection, looking up the threshold associated with the detection's class name
3. If the class is not present in the dictionary, falls back to the configurable default_threshold value
4. Keeps only the detections whose confidence is greater than or equal to the resolved threshold
5. Returns the filtered detections while preserving all original metadata (class ids, masks, keypoints, tracker ids, etc.)

Unlike a single global confidence threshold, this block lets you demand high-confidence predictions for classes that are prone to false positives while keeping a more permissive threshold for classes that are harder to detect. Unlike the generic detections filter, it exposes a purpose-built dictionary input that maps cleanly to a simple {"class_name": threshold} JSON object.

Common Use Cases¶

• Noise-prone classes: Demand very high confidence (e.g. 0.9) for classes that frequently produce false positives, while accepting lower confidence for well-behaved classes
• Hard-to-detect classes: Lower the threshold for classes that the model rarely detects with high confidence so that they are not filtered out entirely
• Production-grade filtering: Apply domain-specific thresholds tuned during evaluation so that downstream analytics, alerts, or counting blocks only see detections that meet the project's quality bar
• Multi-class pipelines: Combine with object detection models that predict many classes at once when a single global confidence threshold is too coarse

Connecting to Other Blocks¶

The filtered predictions from this block can be connected to:

• Visualization blocks (Bounding Box Visualization, Label Visualization, Polygon Visualization) to render only detections that cleared their per-class threshold
• Counting and analytics blocks (Line Counter, Time in Zone, Velocity) so that metrics reflect only high-quality detections
• Tracking blocks (Byte Tracker) so that tracker associations are not polluted by low-confidence noise
• Storage or sink blocks (Roboflow Dataset Upload, Webhook Sink, CSV Formatter) so that only detections meeting the quality bar are persisted or transmitted
• Downstream transformation blocks (Dynamic Crop, Detection Offset) for subsequent processing on the filtered subset

Type identifier¶

Use the following identifier in step "type" field: roboflow_core/per_class_confidence_filter@v1to add the block as as step in your workflow.

Properties¶

The Refs column marks possibility to parametrise the property with dynamic values available in workflow runtime. See Bindings for more info.

Available Connections¶

Input and Output Bindings¶

The available connections depend on its binding kinds. Check what binding kinds Per-Class Confidence Filter in version v1 has.

{
    "name": "<your_step_name_here>",
    "type": "roboflow_core/per_class_confidence_filter@v1",
    "predictions": "$steps.object_detection_model.predictions",
    "class_thresholds": {
        "car": 0.5,
        "person": 0.98
    },
    "default_threshold": 0.3
}