Status: Needs Review

This page has not been reviewed for accuracy and completeness. Content may be outdated or contain errors.

Plugin Nodes¶

Externally contributed nodes loaded through the plugin system. Each plugin ships its own Git repository and is loaded at runtime via a YAML manifest (see Plugin System Overview).

Plugin	Use Case	Nodes
AdaCLIP	Vision-language anomaly detection	`AdaCLIPDetector`
Ultralytics	YOLO26 object detection	`YOLOPreprocess`, `YOLO26Detection`, `YOLOPostprocess`
DeepEIoU	Multi-object tracking with optional ReID	`DeepEIoUTrack`, `OSNetExtractor`, `ResNetExtractor`
SAM3	Video object segmentation and tracking	`SAM3TrackerInference`, `SAM3TextPropagation`, `SAM3BboxPropagation`, `SAM3PointPropagation`, `SAM3MaskPropagation`, `SAM3SegmentEverything`
TrackEval	Tracking evaluation metrics	`HOTAMetricNode`, `CLEARMetricNode`, `IdentityMetricNode`

AdaCLIP¶

Repository: cuvis-ai-adaclip · Manifest: configs/plugins/adaclip.yaml · Tag: v0.1.2

Adapts CLIP (Contrastive Language-Image Pre-training) for hyperspectral anomaly detection. AdaCLIP performs zero-shot visual anomaly scoring using vision-language alignment, making it powerful for detecting anomalies without extensive task-specific training.

Node	Description
`AdaCLIPDetector`	Zero-shot anomaly detection using CLIP-adapted vision-language features

Tutorial: AdaCLIP Workflow

Ultralytics¶

Repository: cuvis-ai-ultralytics · Manifest: configs/plugins/ultralytics.yaml · Tag: v0.1.0

Wraps Ultralytics YOLO for frame-level object detection. Provides preprocessing, inference, and postprocessing stages that slot into any tracking pipeline as the detection front-end.

Node	Description
`YOLOPreprocess`	Resizes and normalizes RGB images for YOLO26 inference
`YOLO26Detection`	Runs YOLO26 object detection on prepared frames
`YOLOPostprocess`	Applies NMS filtering, confidence thresholding, and class filtering

Example: examples/object_tracking/bytetrack/, examples/object_tracking/deepeiou/

DeepEIoU¶

Repository: cuvis-ai-deepeiou · Manifest: configs/plugins/deepeiou.yaml · Tag: v0.1.0

Multi-object tracker using Deep EIoU (Efficient IoU) distance with optional appearance-based association via ReID embeddings. Tracks single or multiple COCO classes and outputs COCO-format detection and tracking JSON.

Node	Description
`DeepEIoUTrack`	Multi-object tracker using EIoU distance metric with optional appearance matching
`OSNetExtractor`	Lightweight ReID feature extractor using OSNet backbone
`ResNetExtractor`	ReID feature extractor using ResNet backbone

Example: examples/object_tracking/deepeiou/

SAM3¶

Repository: cuvis-ai-sam3 · Manifest: configs/plugins/sam3.yaml · Tag: v0.1.3

Video object segmentation and tracking using SAM 2.1 (Segment Anything Model 2.1). Supports multiple prompt strategies — text, bounding box, point, mask, or fully automatic — and works on both CU3S hyperspectral and RGB video inputs.

Node	Description
`SAM3TrackerInference`	Streaming SAM3 tracker for video sequences
`SAM3TextPropagation`	Text-prompted object tracking (e.g. "person", "dog")
`SAM3BboxPropagation`	Bounding-box-prompted object tracking from COCO detection JSON
`SAM3PointPropagation`	Single-point-prompted object tracking from detection JSON
`SAM3MaskPropagation`	Mask-prompted object tracking from segmentation JSON
`SAM3SegmentEverything`	Automatic segmentation of all objects per frame (prompt-free)

Prompt contract¶

Bbox and mask propagation use a scheduled prompt syntax:

--prompt <object_id:detection_id@frame_id>

object_id — tracker-side object ID maintained by SAM3
detection_id — annotation ID copied from the supplied detection or tracking JSON
frame_id — frame where the prompt is injected

Frames before the first scheduled prompt emit empty tracking outputs.

Example: examples/object_tracking/sam3/

TrackEval¶

Repository: cuvis-ai-trackeval · Manifest: configs/plugins/trackeval.yaml · Tag: v0.1.0

Tracking evaluation metrics from TrackEval that compare predicted tracking JSON against ground-truth in COCO bbox format. Requires aligned frame counts between GT and prediction files.

Node	Description
`HOTAMetricNode`	HOTA — Higher Order Tracking Accuracy, reports DetA, AssA, LocA components
`CLEARMetricNode`	CLEAR MOT metrics — reports MOTA, MOTP, FP, FN, ID switches
`IdentityMetricNode`	Identity metrics — reports IDF1, IDP, IDR

Example: examples/object_tracking/trackeval/

Loading Plugins¶

Load a single plugin manifest:

from cuvis_ai_core.utils.node_registry import NodeRegistry

registry = NodeRegistry()
registry.load_plugins("configs/plugins/sam3.yaml")

Or load all official plugins at once from the central registry:

uv run restore-pipeline \
    --pipeline-path configs/pipeline/my_pipeline.yaml \
    --plugins-path configs/plugins/registry.yaml

Combined Manifest¶

Create a custom manifest when you need multiple plugins in one file:

plugins:
  ultralytics:
    repo: "https://github.com/cubert-hyperspectral/cuvis-ai-ultralytics.git"
    tag: "v0.1.0"
    provides:
      - cuvis_ai_ultralytics.node.YOLOPreprocess
      - cuvis_ai_ultralytics.node.YOLO26Detection
      - cuvis_ai_ultralytics.node.YOLOPostprocess

  deepeiou:
    repo: "https://github.com/cubert-hyperspectral/cuvis-ai-deepeiou.git"
    tag: "v0.1.0"
    provides:
      - cuvis_ai_deepeiou.node.DeepEIoUTrack
      - cuvis_ai_deepeiou.node.OSNetExtractor
      - cuvis_ai_deepeiou.node.ResNetExtractor

  trackeval:
    repo: "https://github.com/cubert-hyperspectral/cuvis-ai-trackeval.git"
    tag: "v0.1.0"
    provides:
      - cuvis_ai_trackeval.node.HOTAMetricNode
      - cuvis_ai_trackeval.node.CLEARMetricNode
      - cuvis_ai_trackeval.node.IdentityMetricNode

  sam3:
    repo: "https://github.com/cubert-hyperspectral/cuvis-ai-sam3.git"
    tag: "v0.1.3"
    provides:
      - cuvis_ai_sam3.node.SAM3TrackerInference
      - cuvis_ai_sam3.node.SAM3TextPropagation
      - cuvis_ai_sam3.node.SAM3BboxPropagation
      - cuvis_ai_sam3.node.SAM3PointPropagation
      - cuvis_ai_sam3.node.SAM3MaskPropagation
      - cuvis_ai_sam3.node.SAM3SegmentEverything

Then load it with:

registry = NodeRegistry()
registry.load_plugins("plugins.yaml")

Troubleshooting¶

If a local plugin path fails, verify the path is correct relative to the manifest file.
If a Git plugin fails, verify the tag exists and the repo is accessible from the current environment.
If a node cannot be found after loading, check that the class path appears in provides:.

See Plugin System Overview for architecture details and Plugin Development Guide for creating new plugins.