Roboflow Plugin

Object detection using Roboflow's hosted inference API for vision-agents.

Installation

uv add "vision-agents[roboflow]"
# or directly
uv add vision-agents-plugins-roboflow

Quick Start

from vision_agents.plugins import roboflow
from vision_agents.core import Agent

# A Roboflow processor for cloud inference and
# pre-trained models from Roboflow Universe https://universe.roboflow.com/.
processor = roboflow.RoboflowCloudDetectionProcessor(
    api_key="your_api_key",  # or set ROBOFLOW_API_KEY env var
    api_url="https://detect.roboflow.com",  # or set ROBOFLOW_API_URL env var
    model_id="football-players-detection-3zvbc/20",
    classes=["player"],
    conf_threshold=0.5,
    fps=3,
)

# You can also use a Roboflow processor with local inference and RF-DETR models.
processor = roboflow.RoboflowLocalDetectionProcessor(
    model_id="rfdetr-seg-preview",
    conf_threshold=0.5,
    classes=["person"],
    fps=3,
    # You can pass a custom model as "model" parameter here.
    # The model must be an instance of `rfdetr.RFDETR()` class.
    # model=MyRF_DETRModel()
)

# Use in an agent
agent = Agent(
    processors=[processor],
    llm=your_llm,
    # ... other components
)

Full Example

See example/roboflow_example.py for a complete working example with a video call agent that uses Roboflow detection.

RoboflowCloudDetectionProcessor Configuration

• model_id: Roboflow Universe model id. Example - "football-players-detection-3zvbc/20".
• api_key: Roboflow API key. If not provided, will use ROBOFLOW_API_KEY env variable.
• api_url: Roboflow API url. If not provided, will use ROBOFLOW_API_URL env variable.
• conf_threshold: Confidence threshold for detections (0 - 1.0). Default - 0.5.
• fps: Frame processing rate. Default - 5.
• classes: an optional list of class names to be detected. Example - ["person", "sports ball"] Verify that the classes a supported by the given model. Default - None (all classes are detected). annotate: if True, annotate the detected objects with boxes and labels. Default - True.
• dim_background_factor: how much to dim the background around detected objects from 0 to 1.0. Effective only when annotate=True. Default - 0.0 (no dimming).
• client: an optional custom instance of inference_sdk.InferenceHTTPClient.

RoboflowLocalDetectionProcessor Configuration

• model_id: identifier of the model to be used. Available models are: "rfdetr-base", "rfdetr-large", "rfdetr-nano", "rfdetr-small", "rfdetr-medium", " rfdetr-seg-preview". Default - "rfdetr-seg-preview".

• conf_threshold: Confidence threshold for detections (0 - 1.0). Default - 0.5.

• fps: Frame processing rate. Default - 10.

• classes: optional list of class names to be detected. Example: ["person", "sports ball"] Verify that the classes a supported by the given model. Default - None (all classes are detected).

• annotate: if True, annotate the detected objects with boxes and labels. Default - True.

• dim_background_factor: how much to dim the background around detected objects from 0 to 1.0. Effective only when annotate=True. Default - 0.0 (no dimming).

• model: optional instance of RFDETRModel to be used for detections. Use it provide a model of choosing with custom parameters.

Testing

# Run all tests
pytest plugins/roboflow/tests/ -v

# Run specific tests
pytest plugins/roboflow/tests/test_roboflow.py -v

Dependencies

• vision-agents - Core framework
• numpy>=2.0.0 - Array operations
• rfdetr>=1.3.0 - RF-DETR models for local object detection
• inference-sdk>=0.26.1 - Roboflow SDK for cloud inference

Links

• Roboflow Documentation
• RF-DETR Github
• Roboflow Inference Documentation
• Vision Agents Documentation
• GitHub Repository