DepthAI Nodes

Overview

In DepthAI pipelines, messages are used to send information between nodes. While DepthAI natively supports many message and node types, some use cases need extra host-side components for postprocessing, filtering, visualization, or higher-level pipeline composition.depthai-nodes is an open-source Python library of host-side nodes and message types for DepthAI v3 pipelines. It extends standard DepthAI workflows with parser, utility, and helper nodes that run on the host machine while remaining compatible with the rest of the pipeline.The library is organized into two main modules:

node for host-side pipeline components
message for richer message types used by those components

For more examples and package-level documentation, see the GitHub repository and the API reference.

Installation

depthai-nodes requires Python 3.10 or newer and is intended for depthai v3 pipelines.Install it from PyPI with:

Command Line

1pip install depthai-nodes

If you want to work from source, clone the repository and install it locally:

Command Line

1git clone git@github.com:luxonis/depthai-nodes.git
2cd depthai-nodes
3pip install .

Modules

Node

The node module provides a collection of custom node objects that extend the functionality of standard DepthAI nodes. These nodes are grouped into parser, utility, and helper categories.

Typical workflow

For most AI use cases, ParsingNeuralNetwork is the main entry point. It creates the underlying NeuralNetwork node together with the parser nodes required for the selected model. Model sources can be provided as a Luxonis Models reference, a local NN Archive, or a dai.NNModelDescription.Related helpers include:

HostParsingNeuralNetwork for a host-side version of the same workflow
ParserGenerator when you need to generate parser nodes from an NN Archive

Parsers

Parser nodes are used to interpret the outputs of AI models. Their primary purpose is to encapsulate and abstract away postprocessing logic, allowing users to interact directly with clean, structured outputs. Depending on the parser, outputs can be standard DepthAI messages such as dai.ImgDetections, dai.TrackedFeatures, or dai.ImgFrame, or custom depthai_nodes.message types.Below is an overview of the available parser nodes, along with their corresponding tasks and usage examples. We list only a small fraction of examples as many more are available on Luxonis Models. For detailed information on each individual parser, please refer to the Parser Node documentation.

Parser	Task	Usage Example
`YOLOExtendedParser`	Object Detection	YOLOv8 Instance Segmentation Nano, YOLOv8 Nano Pose Estimation
`YuNetParser`	Object Detection	YuNet
`SCRFDParser`	Object Detection	SCRFD Person detection
`MPPalmDetectionParser`	Object Detection	MediaPipe Palm Detection
`PPTextDetectionParser`	Object Detection	Paddle Text Detection
`ClassificationParser`	Classification	EfficientNet-Lite
`ClassificationSequenceParser`	Classification	Paddle Text Recognition
`SegmentationParser`	Segmentation	DeepLab-V3-Plus
`FastSAMParser`	Segmentation	FastSAM s
`KeypointParser`	Keypoints Detection	MediaPipe Hand Landmarker
`SuperAnimalParser`	Keypoints Detection	SuperAnimal Landmarker
`HRNetParser`	Keypoints Detection	Lite-HRNet
`XFeatMonoParser`	Feature Matching	XFeat
`XFeatStereoParser`	Feature Matching	XFeat
`LaneDetectionParser`	Other	Ultra Fast Lane Detection
`MLSDParser`	Other	M-LSD
`EmbeddingsParser`	Other	ArcFace
`RegressionParser`	Other	Head pose estimation
`MapOutputParser`	Other	Depth Anything V2
`ImageOutputParser`	Other	Zero-DCE

Utility & Helpers

This group includes nodes designed to assist with common tasks such as data processing, filtering, and pipeline management. It also contains abstract base nodes that serve as templates for building custom node implementations.Below is a brief overview of the available utility and helper nodes, along with their corresponding tasks and usage explanations. You can find more information about each individual node in the Utility & Helper Nodes documentation.

Node	Task	Purpose
`BaseHostNode`	Abstract Class	Host node base class
`BaseThreadedHostNode`	Abstract Class	Threaded host node base class
`ApplyColormap`	Image Processing	Applies a colormap
`ApplyDepthColormap`	Image Processing	Applies percentile-normalized depth or disparity coloring
`DepthMerger`	Image Processing	Merges depth and detections
`FrameCropper`	Image Processing	Crops and resizes image regions from detections or upstream manip-config groups
`ImgFrameOverlay`	Image Processing	Overlays two images
`Tiling`	Image Processing	Produces tiled `dai.ImageManipConfig` groups for downstream cropping pipelines
`ExtendedNeuralNetwork`	Neural Network Processing	Convenience wrapper around `dai.node.NeuralNetwork` for model-zoo and NN archive sources
`ParsingNeuralNetwork`	Neural Network Processing	Creates a `NeuralNetwork` node and relevant parser(s) for the given model
`HostParsingNeuralNetwork`	Neural Network Processing	Host-side `ParsingNeuralNetwork` implementation
`ParserGenerator`	Neural Network Processing	Generates parsers from the given NN archive
`CoordinatesMapper`	Detection and Filtering	Remaps coordinates between image-transformation spaces for supported messages
`HostSpatialsCalc`	Detection and Filtering	Computes host-side spatial coordinates using depth data and calibration information
`ImgDetectionsFilter`	Detection and Filtering	Filters detections message
`InstanceToSemanticMask`	Detection and Filtering	Converts instance-id masks into semantic masks
`GatherData`	Data Management	Gathers messages into one
`MessageCollector`	Data Management	Collects same-timestamp messages from a stream into a single collection
`SnapsUploader`	Cloud Integration	Uploads `SnapData` messages to the Luxonis Hub Events API

Message

The message module provides a set of custom message types that extend the standard DepthAI messages. These custom types make it easier to work with the outputs of various AI models and integrate them into visualization or processing pipelines.Here is a quick overview of the message types currently implemented in the library. For field-level details, see the Message Type documentation and the API reference.

Message Type	Stores
`Classifications`	Object class labels with probability scores
`Cluster`	A group of related points
`Clusters`	Multiple `Cluster` objects
`Collection`	A typed collection of messages or other same-type items
`GatheredData`	Reference data together with the gathered messages for that reference
`Keypoints`	Keypoints together with optional skeleton edges
`Line`	Line coordinates with confidence
`Lines`	Multiple `Line` objects
`Map2D`	2D float array
`Prediction`	Single float value
`Predictions`	Multiple `Prediction` objects
`SegmentationMask`	2D integer array
`SnapData`	Single snap event with files, tags, and extra metadata for Luxonis Hub upload

ON THIS PAGE

DepthAI NodesView as Markdown