Feature Request: Support for Multiple ONNX Models in Detection Pipeline

[emrepolat7]

Problem Description

I have two specialized ONNX models that I need to use together:

1. COCO-trained model: Detects persons accurately
2. WiderFace-trained model (or any custom model that natively detects faces): Detects faces accurately but doesn't detect persons

For face recognition to work properly in Frigate 0.16, the system needs to:

1. First detect a person (required by Frigate's face recognition workflow)
2. Then detect faces within the person detection area

Currently, while Frigate supports multiple detectors, each detector can only use a single model. This limits the ability to combine specialized models within a single detection pipeline.

Proposed Solution

Add support for multiple ONNX models in a pipeline configuration, such as:

detectors:
primary:
type: onnx
model:
path: /models/coco_person_detector.onnx
secondary:
type: onnx
model:
path: /models/widerface_detector.onnx
depends_on: primary
roi_from: person # Only run face detection within person detections

Alternative Workarounds Considered

• Model merging: Too complex and requires retraining
• External processing: Adds latency and complexity
• Custom detector plugin: Requires significant development effort

Use Cases
This would benefit users who have:

Specialized models for different object types
Models optimized for specific scenarios (indoor/outdoor, day/night)
Legacy models that work well for specific use cases but lack comprehensive object detection

Expected Behavior
The detection pipeline would:

Run the primary detector (person detection)
For each person detection, run secondary detector(s) within that ROI
Combine results into a unified detection output
Pass combined results to face recognition system

This feature would significantly improve detection accuracy by allowing users to leverage specialized models for different detection tasks.