Detector
Runs on full frames and produces bounding box detections. At least one primary detector (without operate_on) is required in every pipeline.
Pose estimation
Detects human keypoints (17 COCO joints) in addition to bounding boxes. The pipeline draws skeleton overlays on the output video automatically.
Instance segmentation
Detects objects and produces per-instance segmentation masks. Mask overlays are drawn automatically on the output video.
Classifier (model chaining)
Classifiers run on objects detected by a parent model. They must reference a previously defined model using operate_on.
Triton detector
For models served by NVIDIA Triton Inference Server. Uses gRPC to communicate with the Triton server.
Triton classifier
Classification via Triton Inference Server. Must chain from a parent model using operate_on.
Tracker
An object tracker can be placed anywhere in the models list using type: "tracker". Its position determines where tracking occurs in the inference chain (e.g., between a detector and a classifier).
Active classes
Use active_classes on a detector, pose, or segmentation model to only detect specific classes, ignoring all others. This is useful when your model outputs many classes but you only care about a few.