Detections Consensus¶

Combine detection predictions from multiple models using a majority vote consensus strategy, merging overlapping detections that receive sufficient votes from different models and aggregating their properties (confidence scores, bounding boxes, masks) into unified consensus detections with improved accuracy and reliability.

How This Block Works¶

This block fuses predictions from multiple detection models by requiring agreement (consensus) among models before accepting detections. The block:

1. Takes detection predictions from multiple model sources (object detection, instance segmentation, or keypoint detection) as input
2. Matches detections from different models that overlap spatially by calculating Intersection over Union (IoU) between bounding boxes
3. Compares overlapping detections against an IoU threshold to determine if they represent the same object
4. Counts "votes" for each detection by finding matching detections from other models (subject to class-awareness if enabled)
5. Requires a minimum number of votes (required_votes) before accepting a detection as part of the consensus output
6. Aggregates properties of matching detections using configurable modes:
7. Confidence aggregation: Combines confidence scores using average, max, or min
8. Coordinates aggregation: Merges bounding boxes using average (mean coordinates), max (largest box), or min (smallest box)
9. Mask aggregation (for instance segmentation): Combines masks using union, intersection, max, or min
10. Class selection: Chooses class name based on majority vote (average), highest confidence (max), or lowest confidence (min)
11. Filters detections based on optional criteria (specific classes to consider, minimum confidence threshold)
12. Determines object presence by checking if the required number of objects (per class or total) are present in consensus results
13. Returns merged consensus detections, object presence indicators, and presence confidence scores

The block enables class-aware or class-agnostic matching: when class_aware is true, only detections with matching class names are considered for voting; when false, any overlapping detections (regardless of class) contribute votes. The consensus mechanism helps reduce false positives (detections seen by only one model) and improves reliability by requiring multiple models to agree on object presence. Aggregation modes allow flexibility in how overlapping detections are combined, balancing between conservative (intersection, min) and inclusive (union, max) strategies.

Common Use Cases¶

• Multi-Model Ensemble: Combine predictions from multiple specialized models (e.g., one optimized for people, another for vehicles) to improve overall detection accuracy, leveraging strengths of different models while filtering out detections that only one model sees
• Reducing False Positives: Require consensus from multiple models before accepting detections (e.g., require 2 out of 3 models to detect an object), reducing false positives by filtering out detections seen by only one model
• Improving Detection Reliability: Use majority voting to increase confidence in detections (e.g., merge overlapping detections from 3 models, keeping only those with 2+ votes), ensuring only high-confidence, multi-model-agreed detections are retained
• Object Presence Detection: Determine if specific objects are present based on consensus (e.g., check if at least 2 "person" detections exist across models, use aggregated confidence to determine presence), enabling robust object presence checking with configurable thresholds
• Class-Specific Consensus: Apply different consensus requirements per class (e.g., require 3 votes for "car" but only 2 for "person"), allowing stricter criteria for critical objects while being more lenient for common detections
• Specialized Model Fusion: Combine general-purpose and specialized models (e.g., general object detector + specialized license plate detector), creating a unified detection system that benefits from both broad coverage and specific expertise

Connecting to Other Blocks¶

The consensus predictions from this block can be connected to:

• Multiple detection model blocks (e.g., Object Detection Model, Instance Segmentation Model) to receive predictions from different models that are fused into consensus detections based on majority voting and spatial overlap matching
• Visualization blocks (e.g., Bounding Box Visualization, Polygon Visualization, Label Visualization) to display the merged consensus detections, showing unified results from multiple models with improved accuracy
• Counting and analytics blocks (e.g., Line Counter, Time in Zone, Velocity) to count or analyze consensus detections, providing more reliable metrics based on multi-model agreement
• Data storage blocks (e.g., Local File Sink, CSV Formatter, Roboflow Dataset Upload, Webhook Sink) to save or transmit consensus detection results, storing fused predictions that represent multi-model agreement
• Flow control blocks (e.g., Continue If) to conditionally trigger downstream processing based on object_present indicators or presence_confidence scores, enabling workflows that respond to multi-model consensus on object presence
• Filtering blocks (e.g., Detections Filter) to further refine consensus detections based on additional criteria, enabling multi-stage filtering after consensus fusion

Type identifier¶

Use the following identifier in step "type" field: roboflow_core/detections_consensus@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 Detections Consensus in version v1 has.

{
    "name": "<your_step_name_here>",
    "type": "roboflow_core/detections_consensus@v1",
    "predictions_batches": [
        "$steps.a.predictions",
        "$steps.b.predictions"
    ],
    "required_votes": 2,
    "class_aware": true,
    "iou_threshold": 0.3,
    "confidence": 0.1,
    "classes_to_consider": [
        "a",
        "b"
    ],
    "required_objects": 3,
    "presence_confidence_aggregation": "max",
    "detections_merge_confidence_aggregation": "min",
    "detections_merge_coordinates_aggregation": "min",
    "detections_merge_mask_aggregation": "union"
}