Building Hybrid Pipelines

The section discusses how to pack multiple pipelines in a single pipeline, separating their compute spaces with custom primary ROI. In Savant, a developer can pack multiple sub-pipelines into a single pipeline, working either on all frames or conditionally based on ROI.

Unconditional Processing

By default, the primary model analyzes the whole frame; however, under the hood, Savant creates the default top-level object covering the whole frame; thus, the models without specified input constraints can analyze it. It allows placing multiple primary models one after another and then their secondary models. The only requirement is non-overlapping unit names to avoid object collisions.

Normally, there is no difference how to place units if there are no cross dependencies between units; thus, the ordering is important only between elements of sub-pipelines.


consider placing sub-pipelines in element groups: it helps to develop and debug them independently.

Conditional Processing

When you need to process frames conditionally, based on per-stream information, e.g., handle cam-1 with a car processing sub-pipeline and cam-2 with a person processing sub-pipeline, a developer must place a special ROI-modifying custom pyfunc before other pipeline elements.

That pyfunc must modify ROI based on source-id or other knowledge like per-frame attributes:

  • when source-id is unknown it can be deleted to ensure the frame is not processed;

  • when source-id is known and relates to the car processing sub-pipeline, set it to car.roi;

  • when source-id is known and relates to the person processing sub-pipeline, set it to person.roi;

The primary models must accept corresponding ROIs rather than work on default ROI:

element: nvinfer@detector
  name: CarDetector
      object: car.roi
element: nvinfer@detector
  name: PersonDetector
      object: person.roi

Pros & Cons Of Hybrid Pipelines


  • easier to maintain deployments;

  • efficient processing;

  • easier to route video streams (no stream duplication is needed);


  • more difficult to develop and troubleshoot, consider element groups;

  • increases end-to-end delay;

  • more difficult to plan compute resources when real-time processing is required.