Skip to main content

Adding a New Operation

EagleEye has secondary operations (single-file processing) and main operations (a definition plus heavier implementation code). Both extend OperationInstance. Operation source is added in the repository; the WebUI has no Custom Ops source editor.

1. Add the class​

For a simple secondary operation, create src/secondary_operations/my_transform.py:

from typing import Any
from src.main_operations.definitions.base.base_class import OperationInstance

class MyTransform(OperationInstance):
    def __init__(self, min_confidence: float = 0.5) -> None:
        self.min_confidence = min_confidence

    def run(self, input_data: Any) -> Any:
        if input_data is None:
            return None
        return [item for item in input_data if item["confidence"] >= self.min_confidence]

For inference or shared implementation code, put a *Definition class in src/main_operations/definitions/ and implementation code below src/main_operations/modules/. Constructors can request device_registry and model_library; pipeline construction injects both. Validate the exact canonical device ID (cpu, cuda:N, or mx3:N) and resolve its managed artifact rather than using the removed ComputePool API.

class MyModelDefinition(OperationInstance):
    def __init__(self, model_id, device_id, device_registry, model_library):
        device_registry.get(device_id)
        artifact = model_library.resolve_artifact(model_id, device_id)
        # Load artifact.path in this operation's runtime.

2. Add the config definition​

Create a matching JSON file:

  • secondary: src/secondary_operations/config_data/my_transform_config_def.json
  • main: src/main_operations/definitions/config_data/my_model_config_def.json

The definition supplies the class name, category, input/output ports, and editable parameters. For managed inference, use ui_hint: "model_library" on model_id and ui_hint: "device_registry" on device_id (with the reciprocal device_param/model_param fields); see an existing object-detection config definition for the complete shape.

3. Add it to a pipeline​

Use the Pipeline tab or edit src/config/pipeline_config.json. action_params contains only configured values, never injected dependencies such as network_table, camera_manager, camera_config_registry, device_registry, model_library, or logger.

4. Restart and verify​

Restart the backend after source/config changes. Confirm the operation appears in the Pipeline palette, its ports connect as expected, and its pipeline starts without construction errors. visualize() may return a BGR frame for the visualization stream; update_config() handles live changes for parameters that do not require restart.