# DepthAI Python API DepthAI Python API can be found on Github [luxonis/depthai-python](https://github.com/luxonis/depthai-python). Below is the reference documentation for the Python API. ### depthai Kind: Package #### beta Kind: Package Experimental APIs ##### node Kind: Module Experimental nodes ###### depthai.beta.node.ClassificationParser(depthai.DeviceNode) Kind: Class ClassificationParser node. Parses the raw output of a classification neural network into a dai::beta::Classifications message with class names and scores sorted in descending order of score. The parser consumes a single output tensor. When the incoming NNData contains exactly one tensor, it is selected automatically; otherwise the output layer name must be configured explicitly or through an NNArchive head. Raw scores are dequantized and flattened; when the model output is not already softmaxed, the parser applies softmax to convert the scores to probabilities. ###### build() Kind: Method ###### getClasses(self) -> list [ str ]: list [ str ] Kind: Method Returns the class names to link with the classification scores. ###### getOutputLayerName(self) -> str: str Kind: Method Returns the name of the model output layer to parse. ###### getSoftmax(self) -> bool: bool Kind: Method Returns whether the model output is treated as already softmaxed. ###### runOnHost(self) -> bool: bool Kind: Method Returns true when this node runs on the host. Host-only pipelines always run the node on the host. ###### setClasses(self, classes: list [ str ]) Kind: Method Sets the class names to link with the classification scores. The class names are expected to be in the same order as the neural network's output. The number of class names must match the number of scores produced by the model. Parameter ``classes``: Vector of class names ###### setNNArchive(self, nnArchive: depthai.NNArchive) Kind: Method Set NNArchive for this Node. The archive must contain exactly one ClassificationParser head; use setNNArchiveHead() to select a specific head from a multi-head archive. Parameter ``nnArchive:``: NNArchive to set ###### setNNArchiveHead(self, head: depthai.nn_archive.v1.Head) Kind: Method Set NNArchive head for this Node. The head must be a ClassificationParser head with exactly one output layer. Parameter ``head:``: NNArchive head to set ###### setOutputLayerName(self, outputLayerName: str) Kind: Method Sets the name of the model output layer to parse. When left empty, the parser selects the tensor automatically if the incoming NNData contains exactly one tensor and fails otherwise. Parameter ``outputLayerName``: Name of the output layer ###### setRunOnHost(self, runOnHost: bool) Kind: Method Select whether the node runs on the host or device. ###### setSoftmax(self, isSoftmax: bool) Kind: Method Sets whether the model output is already softmaxed. When false, the parser applies softmax to convert the raw scores to probabilities. Parameter ``isSoftmax``: True when the model output is already softmaxed ###### input Kind: Property Input NN results with classification data to parse. ###### out Kind: Property Outputs Classifications message with classes and scores sorted in descending order of score. ###### depthai.beta.node.ClassificationSequenceParser(depthai.DeviceNode) Kind: Class ClassificationSequenceParser node. Parses the raw output of a classification sequence neural network into a dai::beta::Classifications message with class names and scores ordered by their position in the sequence. The model predicts the classes multiple times and returns a list of predicted classes, where each item corresponds to the relative step in the sequence. In addition to time series classification, this parser can also be used for text recognition models where words can be interpreted as a sequence of characters (classes). The parser consumes a single output tensor of shape (sequenceLength, nClasses), (1, sequenceLength, nClasses) or (sequenceLength, nClasses, 1). When the incoming NNData contains exactly one tensor, it is selected automatically; otherwise the output layer name must be configured explicitly or through an NNArchive head. Raw scores are dequantized; when the model output is not already softmaxed, the parser applies softmax along each sequence step to convert the scores to probabilities. ###### build() Kind: Method ###### getClasses(self) -> list [ str ]: list [ str ] Kind: Method Returns the class names to link with the per-step classification scores. ###### getConcatenateClasses(self) -> bool: bool Kind: Method Returns whether the remaining classes are concatenated. ###### getIgnoredIndexes(self) -> list [ int ]: list [ int ] Kind: Method Returns the class indexes ignored during classification sequence generation. ###### getOutputLayerName(self) -> str: str Kind: Method Returns the name of the model output layer to parse. ###### getRemoveDuplicates(self) -> bool: bool Kind: Method Returns whether consecutive duplicate classes are removed from the sequence. ###### getSoftmax(self) -> bool: bool Kind: Method Returns whether the model output is treated as already softmaxed. ###### runOnHost(self) -> bool: bool Kind: Method Returns true when this node runs on the host. Host-only pipelines always run the node on the host. ###### setClasses(self, classes: list [ str ]) Kind: Method Sets the class names to link with the per-step classification scores. The class names are expected to be in the same order as the neural network's output. The number of class names must match the number of scores produced by the model at each sequence step. Parameter ``classes``: Vector of class names ###### setConcatenateClasses(self, concatenateClasses: bool) Kind: Method Sets whether the remaining classes are concatenated. Used mostly for text processing. When true and more than one class remains: when all remaining class names are at most one character long, they are joined and split on whitespace into words with a per-word mean score; otherwise all class names are joined into a single string with a " " separator and one mean score. Parameter ``concatenateClasses``: True to concatenate the remaining classes @note Configures startup behavior. Send ClassificationSequenceParserConfig to inputConfig after the pipeline starts. ###### setIgnoredIndexes(self, ignoredIndexes: list [ int ]) Kind: Method Sets the class indexes to ignore during classification sequence generation (e.g. background class, blank space). Sequence steps whose most probable class index is listed here are dropped from the output. Every index must be within [0, nClasses - 1]. Parameter ``ignoredIndexes``: Vector of class indexes to ignore @note Configures startup behavior. Send ClassificationSequenceParserConfig to inputConfig after the pipeline starts. ###### setNNArchive(self, nnArchive: depthai.NNArchive) Kind: Method Set NNArchive for this Node. The archive must contain exactly one ClassificationSequenceParser head; use setNNArchiveHead() to select a specific head from a multi-head archive. Parameter ``nnArchive:``: NNArchive to set ###### setNNArchiveHead(self, head: depthai.nn_archive.v1.Head) Kind: Method Set NNArchive head for this Node. The head must be a ClassificationSequenceParser head with exactly one output layer. Parameter ``head:``: NNArchive head to set ###### setOutputLayerName(self, outputLayerName: str) Kind: Method Sets the name of the model output layer to parse. When left empty, the parser selects the tensor automatically if the incoming NNData contains exactly one tensor and fails otherwise. Parameter ``outputLayerName``: Name of the output layer ###### setRemoveDuplicates(self, removeDuplicates: bool) Kind: Method Sets whether consecutive duplicate classes are removed from the sequence. Only consecutive duplicates are removed; repeated classes separated by other classes are kept. Parameter ``removeDuplicates``: True to remove consecutive duplicates from the sequence @note Configures startup behavior. Send ClassificationSequenceParserConfig to inputConfig after the pipeline starts. ###### setRunOnHost(self, runOnHost: bool) Kind: Method Select whether the node runs on the host or device. ###### setSoftmax(self, isSoftmax: bool) Kind: Method Sets whether the model output is already softmaxed. When false, the parser applies softmax along each sequence step to convert the raw scores to probabilities. Parameter ``isSoftmax``: True when the model output is already softmaxed ###### initialConfig Kind: Property Configuration used until a message is received on inputConfig. ###### input Kind: Property Input NN results with classification sequence data to parse. ###### inputConfig Kind: Property Runtime parser configuration. When synchronized, one configuration is consumed per frame; otherwise all queued configurations are drained and the newest valid one is used. ###### out Kind: Property Outputs Classifications message with classes and scores ordered by their position in the sequence. ###### depthai.beta.node.EmbeddingsParser(depthai.DeviceNode) Kind: Class EmbeddingsParser node. Validates the raw output of an embeddings neural network model head and forwards it unchanged as a dai::NNData message. The parser expects a single output tensor carrying the embedding vector. When the output layer name is left unconfigured, every incoming NNData must contain exactly one tensor; otherwise the message is rejected. The message itself is forwarded without modification, so all tensors, sequence number, timestamps, and image transformation metadata are preserved. ###### build() Kind: Method ###### getOutputLayerName(self) -> str: str Kind: Method Returns the name of the model output layer carrying the embeddings. ###### runOnHost(self) -> bool: bool Kind: Method Returns true when this node runs on the host. Host-only pipelines always run the node on the host. ###### setNNArchive(self, nnArchive: depthai.NNArchive) Kind: Method Set NNArchive for this Node. The archive must contain exactly one EmbeddingsParser head; use setNNArchiveHead() to select a specific head from a multi-head archive. Parameter ``nnArchive:``: NNArchive to set ###### setNNArchiveHead(self, head: depthai.nn_archive.v1.Head) Kind: Method Set NNArchive head for this Node. The head must be an EmbeddingsParser head with exactly one output layer. Parameter ``head:``: NNArchive head to set ###### setOutputLayerName(self, outputLayerName: str) Kind: Method Sets the name of the model output layer carrying the embeddings. When left empty, the parser requires the incoming NNData to contain exactly one tensor and fails otherwise. Parameter ``outputLayerName``: Name of the output layer ###### setRunOnHost(self, runOnHost: bool) Kind: Method Select whether the node runs on the host or device. ###### input Kind: Property Input NN results with embeddings data to validate and forward. ###### out Kind: Property Outputs the unchanged NNData message containing the embeddings output layer. ###### depthai.beta.node.FastSAMParser(depthai.DeviceNode) Kind: Class FastSAMParser node. Parses the output of the FastSAM segmentation model (https://github.com/CASIA-IVA-Lab/FastSAM) into a dai::SegmentationMask message where each pixel holds the index of the instance it belongs to and 255 marks background. The parser consumes the model's YOLO detection outputs (NCHW tensors of shape (1, numClasses + 5, gridH, gridW), sorted by layer name and decoded anchorless with strides 8/16/32), the per-head mask-coefficient outputs (NCHW tensors of shape (1, numPrototypes, gridH, gridW), sorted by layer name) and the prototype masks output (NCHW tensor of shape (1, numPrototypes, protoH, protoW)). The model input size is derived from the first (stride-8) YOLO output's grid times 8; the number of prototypes from the protos tensor's channel count. Boxes pass confidence filtering and non-maximum suppression, boxes within 20 pixels of the image border are snapped to it, and a box overlapping the full image with IoU > 0.9 is replaced by the full-image box. Each kept detection's mask is combined from the prototypes, resized to the model input size with nearest-neighbor interpolation, cropped to its box and binarized with the mask confidence threshold. The prompt selects the emitted instances: "everything" keeps all detections (later, lower-confidence instances overwrite earlier ones on overlapping pixels), "bbox" keeps the single mask with the highest IoU against the prompt bounding box, and "point" combines the masks containing the prompt point (added for point label 1, subtracted for 0). With no detections, a fully-background mask is emitted. ###### build() Kind: Method ###### getBoundingBox(self) -> typing.Annotated [ list [ int ], pybind11_stubgen.typing_ext.FixedSize(4) ]|None: typing.Annotated [ list [ int ], pybind11_stubgen.typing_ext.FixedSize(4) ]|None Kind: Method Returns the prompt bounding box as (x1, y1, x2, y2), or std::nullopt when it is not set. ###### getConfidenceThreshold(self) -> float: float Kind: Method Returns the confidence score threshold for detected objects. ###### getIouThreshold(self) -> float: float Kind: Method Returns the non-maximum suppression overlap threshold. ###### getMaskConfidence(self) -> float: float Kind: Method Returns the mask confidence threshold. ###### getMaskOutputs(self) -> list [ str ]: list [ str ] Kind: Method Returns the names of the model's mask-coefficient output layers. ###### getNumClasses(self) -> int: int Kind: Method Returns the number of classes in the model. ###### getPointLabel(self) -> int|None: int|None Kind: Method Returns the prompt point label, or std::nullopt when it is not set. ###### getPoints(self) -> tuple [ int, int ]|None: tuple [ int, int ]|None Kind: Method Returns the prompt point as (x, y), or std::nullopt when it is not set. ###### getPrompt(self) -> str: str Kind: Method Returns the prompt type. ###### getProtosOutput(self) -> str: str Kind: Method Returns the name of the model's prototype-masks output layer. ###### getYoloOutputs(self) -> list [ str ]: list [ str ] Kind: Method Returns the names of the model's YOLO output layers. ###### runOnHost(self) -> bool: bool Kind: Method Returns true when this node runs on the host. Host-only pipelines always run the node on the host. ###### setBoundingBox(self, bbox: typing.Annotated [ list [ int ], pybind11_stubgen.typing_ext.FixedSize ( 4 ) ]) Kind: Method Sets the prompt bounding box as (x1, y1, x2, y2) in model-input pixels, used by the "bbox" prompt. Unset by default; the "bbox" prompt requires it and its x2 and y2 coordinates must not be 0. Parameter ``bbox``: Bounding box as (x1, y1, x2, y2) @note Configures startup behavior. Send FastSAMParserConfig to inputConfig after the pipeline starts. ###### setConfidenceThreshold(self, threshold: float) Kind: Method Sets the confidence score threshold for detected objects. Detections whose score is strictly greater than the threshold are kept. Defaults to 0.5. Parameter ``threshold``: Confidence score threshold, must be between 0 and 1 @note Configures startup behavior. Send FastSAMParserConfig to inputConfig after the pipeline starts. ###### setIouThreshold(self, iouThreshold: float) Kind: Method Sets the non-maximum suppression overlap threshold. Boxes whose overlap with a kept box is strictly greater than the threshold are suppressed. Defaults to 0.5. Parameter ``iouThreshold``: Overlap threshold, must be between 0 and 1 @note Configures startup behavior. Send FastSAMParserConfig to inputConfig after the pipeline starts. ###### setMaskConfidence(self, maskConfidence: float) Kind: Method Sets the mask confidence threshold used to binarize instance masks. Mask pixels with a sigmoid probability strictly greater than the threshold belong to the instance. Defaults to 0.5. Parameter ``maskConfidence``: Mask confidence threshold, must be between 0 and 1 @note Configures startup behavior. Send FastSAMParserConfig to inputConfig after the pipeline starts. ###### setMaskOutputs(self, maskOutputs: list [ str ]) Kind: Method Sets the names of the model's mask-coefficient output layers. Only names containing "mask" are used, sorted by name and index-aligned with the sorted YOLO output layers; when empty, all layer names of the incoming NNData containing "mask" are used. Defaults to ["output1_masks", "output2_masks", "output3_masks"]. Parameter ``maskOutputs``: Names of the mask output layers ###### setNNArchive(self, nnArchive: depthai.NNArchive) Kind: Method Set NNArchive for this Node. The archive must contain exactly one FastSAMParser head; use setNNArchiveHead() to select a specific head from a multi-head archive. Parameter ``nnArchive:``: NNArchive to set ###### setNNArchiveHead(self, head: depthai.nn_archive.v1.Head) Kind: Method Set NNArchive head for this Node. The head's output layer names containing "_yolo" configure the YOLO output layers and those containing "_masks" the mask output layers (each only when at least one matches). The confidence threshold, number of classes, NMS threshold, mask confidence, prompt, points, point label and bounding box are read from the head metadata when present. Parameter ``head:``: NNArchive head to set ###### setNumClasses(self, numClasses: int) Kind: Method Sets the number of classes in the model. The YOLO output tensors must have numClasses + 5 channels. Defaults to 1. Parameter ``numClasses``: Number of classes, must be greater than 0 ###### setPointLabel(self, pointLabel: int) Kind: Method Sets the prompt point label, used by the "point" prompt: 1 adds the instance masks containing the point, 0 subtracts them. Unset by default; the "point" prompt requires it. Parameter ``pointLabel``: Point label @note Configures startup behavior. Send FastSAMParserConfig to inputConfig after the pipeline starts. ###### setPoints(self, x: int, y: int) Kind: Method Sets the prompt point as (x, y) in model-input pixels, used by the "point" prompt. Unset by default; the "point" prompt requires it. Parameter ``x``: Point x coordinate Parameter ``y``: Point y coordinate @note Configures startup behavior. Send FastSAMParserConfig to inputConfig after the pipeline starts. ###### setPrompt(self, prompt: str) Kind: Method Sets the prompt type: "everything" emits every detected instance, "bbox" the single instance mask with the highest IoU against the prompt bounding box (see setBoundingBox()), and "point" the combination of the instance masks containing the prompt point (see setPoints() and setPointLabel()). Defaults to "everything". Parameter ``prompt``: Prompt type, one of "everything", "bbox" or "point" @note Configures startup behavior. Send FastSAMParserConfig to inputConfig after the pipeline starts. ###### setProtosOutput(self, protosOutput: str) Kind: Method Sets the name of the model's prototype-masks output layer; when empty, "protos_output" is used. Defaults to "protos_output". Parameter ``protosOutput``: Name of the protos output layer ###### setRunOnHost(self, runOnHost: bool) Kind: Method Select whether the node runs on the host or device. ###### setYoloOutputs(self, yoloOutputs: list [ str ]) Kind: Method Sets the names of the model's YOLO output layers. The layers are processed sorted by name, so the stride-8 head must come first in sort order. Defaults to ["output1_yolov8", "output2_yolov8", "output3_yolov8"]. Parameter ``yoloOutputs``: Names of the YOLO output layers ###### initialConfig Kind: Property Configuration used until a message is received on inputConfig. ###### input Kind: Property Input NN results with FastSAM data to parse. ###### inputConfig Kind: Property Runtime parser configuration. When synchronized, one configuration is consumed per frame; otherwise all queued configurations are drained and the newest valid one is used. ###### out Kind: Property Outputs SegmentationMask message with the resulting segmentation masks given the prompt. ###### depthai.beta.node.HRNetParser(depthai.DeviceNode) Kind: Class HRNetParser node. Parses the heatmap output of an HRNet pose estimation neural network into a dai::beta::Keypoints message. The decoding is inspired by https://github.com/ibaiGorordo/ONNX-HRNET-Human-Pose-Estimation. The parser consumes a single output tensor. When the incoming NNData contains exactly one tensor, it is selected automatically; otherwise the output layer name must be configured explicitly or through an NNArchive head. The tensor is read in NCHW orientation regardless of its stored order; after squeezing a leading batch dimension of 1 it must be a 3D tensor of shape (numKeypoints, height, width). The number of keypoints and the heatmap size are derived from the tensor shape. Per heatmap, the keypoint is the position of the maximum value normalized by the heatmap size and the keypoint's score is the maximum value clipped to [0, 1]. Keypoints with a score below the score threshold are dropped and the skeleton edges are remapped to the kept keypoints. ###### build() Kind: Method ###### getEdges(self) -> list [ typing.Annotated [ list [ int ], pybind11_stubgen.typing_ext.FixedSize(2) ] ]: list [ typing.Annotated [ list [ int ], pybind11_stubgen.typing_ext.FixedSize(2) ] ] Kind: Method Returns the skeleton edges as pairs of keypoint indices. ###### getLabelNames(self) -> list [ str ]: list [ str ] Kind: Method Returns the label names for the keypoints. ###### getOutputLayerName(self) -> str: str Kind: Method Returns the name of the model output layer to parse. ###### getScoreThreshold(self) -> float: float Kind: Method Returns the confidence score threshold for detected keypoints. ###### runOnHost(self) -> bool: bool Kind: Method Returns true when this node runs on the host. Host-only pipelines always run the node on the host. ###### setEdges(self, edges: list [ typing.Annotated [ list [ int ], pybind11_stubgen.typing_ext.FixedSize ( 2 ) ] ]) Kind: Method Sets the skeleton edges as pairs of keypoint indices used for visualizing the skeleton. Example: {{0, 1}, {1, 2}, {2, 3}, {3, 0}} connects keypoint 0 to keypoint 1, keypoint 1 to keypoint 2, etc. Parameter ``edges``: Vector of keypoint index pairs ###### setLabelNames(self, labelNames: list [ str ]) Kind: Method Sets the label names for the keypoints, indexed by keypoint index. Parameter ``labelNames``: Vector of label names ###### setNNArchive(self, nnArchive: depthai.NNArchive) Kind: Method Set NNArchive for this Node. The archive must contain exactly one HRNetParser head; use setNNArchiveHead() to select a specific head from a multi-head archive. Parameter ``nnArchive:``: NNArchive to set ###### setNNArchiveHead(self, head: depthai.nn_archive.v1.Head) Kind: Method Set NNArchive head for this Node. The head must be an HRNetParser head with exactly one output layer. Parameter ``head:``: NNArchive head to set ###### setOutputLayerName(self, outputLayerName: str) Kind: Method Sets the name of the model output layer to parse. When left empty, the parser selects the tensor automatically if the incoming NNData contains exactly one tensor and fails otherwise. Parameter ``outputLayerName``: Name of the output layer ###### setRunOnHost(self, runOnHost: bool) Kind: Method Select whether the node runs on the host or device. ###### setScoreThreshold(self, threshold: float) Kind: Method Sets the confidence score threshold for detected keypoints. Keypoints with a score strictly below the threshold are dropped. Parameter ``threshold``: Confidence score threshold, must be between 0 and 1 @note Configures startup behavior. Send HRNetParserConfig to inputConfig after the pipeline starts. ###### initialConfig Kind: Property Configuration used until a message is received on inputConfig. ###### input Kind: Property Input NN results with heatmaps data to parse. ###### inputConfig Kind: Property Runtime parser configuration. When synchronized, one configuration is consumed per frame; otherwise all queued configurations are drained and the newest valid one is used. ###### out Kind: Property Outputs Keypoints message with the detected body keypoints. ###### depthai.beta.node.ImageOutputParser(depthai.DeviceNode) Kind: Class ImageOutputParser node. Parses the output of image-to-image models (e.g. DnCNN3, zero-dce) where the output is a modified image (denoised, enhanced etc.) into a dai::ImgFrame message. The parser consumes a single output tensor. When the incoming NNData contains exactly one tensor, it is selected automatically; otherwise the output layer name must be configured explicitly or through an NNArchive head. The tensor is read in its stored order; after squeezing a leading batch dimension of 1 it must be a 3D image tensor in CHW or HWC orientation, with the channel dimension equal to 1 (grayscale) or 3 (color). All dimensions are derived from the runtime tensor descriptor. The values are min-max normalized and scaled to the [0, 255] 8-bit range. A grayscale image is emitted as a GRAY8 frame. A color image is emitted as a BGR888p frame when the pipeline's default device platform is RVC2 and as a BGR888i frame otherwise, including in a device-less pipeline. The model output is treated as RGB and converted to BGR unless the BGR-output flag marks it as already BGR. ###### build() Kind: Method ###### getBGROutput(self) -> bool: bool Kind: Method Returns the flag indicating whether the model output image is in BGR (Blue- Green-Red) channel order. ###### getOutputLayerName(self) -> str: str Kind: Method Returns the name of the model output layer to parse. ###### runOnHost(self) -> bool: bool Kind: Method Returns true when this node runs on the host. Host-only pipelines always run the node on the host. ###### setBGROutput(self, outputIsBGR: bool = True) Kind: Method Sets the flag indicating whether the model output image is in BGR (Blue-Green- Red) channel order. When false (the default), a color model output is treated as RGB and its channels are swapped to BGR before being emitted. Parameter ``outputIsBGR``: True when the model output image is already BGR, defaults to true ###### setNNArchive(self, nnArchive: depthai.NNArchive) Kind: Method Set NNArchive for this Node. The archive must contain exactly one ImageOutputParser head; use setNNArchiveHead() to select a specific head from a multi-head archive. Parameter ``nnArchive:``: NNArchive to set ###### setNNArchiveHead(self, head: depthai.nn_archive.v1.Head) Kind: Method Set NNArchive head for this Node. The head must be an ImageOutputParser head with exactly one output layer. Parameter ``head:``: NNArchive head to set ###### setOutputLayerName(self, outputLayerName: str) Kind: Method Sets the name of the model output layer to parse. When left empty, the parser selects the tensor automatically if the incoming NNData contains exactly one tensor and fails otherwise. Parameter ``outputLayerName``: Name of the output layer ###### setRunOnHost(self, runOnHost: bool) Kind: Method Select whether the node runs on the host or device. ###### input Kind: Property Input NN results with image tensor data to parse. ###### out Kind: Property Outputs ImgFrame message with the model output image, e.g. a denoised or enhanced image. ###### depthai.beta.node.ImgDetectionsFilter(depthai.DeviceNode) Kind: Class Experimental node for filtering image detections. ###### runOnHost(self) -> bool: bool Kind: Method Returns true when this node runs on the host. Host-only pipelines always run the node on the host. ###### setRunOnHost(self, runOnHost: bool) Kind: Method Select whether the node runs on the host or device. ###### initialConfig Kind: Property Configuration used until a message is received on inputConfig. The default configuration forwards detections unchanged. ###### input Kind: Property Image detections to filter. ###### inputConfig Kind: Property Runtime filter configuration. The most recently received configuration is reused for subsequent detection messages. ###### output Kind: Property Filtered image detections. ###### depthai.beta.node.KeypointParser(depthai.DeviceNode) Kind: Class KeypointParser node. Parses the raw output of a 2D or 3D keypoints neural network into a dai::beta::Keypoints message. The parser consumes a single output tensor. When the incoming NNData contains exactly one tensor, it is selected automatically; otherwise the output layer name must be configured explicitly or through an NNArchive head. The number of keypoints must be configured before the pipeline starts. The number of coordinates per keypoint (2 or 3) is derived from the tensor size and the configured number of keypoints. Keypoint coordinates are divided by the configured scale factor and clipped to [0, 1]. ###### build() Kind: Method ###### getEdges(self) -> list [ typing.Annotated [ list [ int ], pybind11_stubgen.typing_ext.FixedSize(2) ] ]: list [ typing.Annotated [ list [ int ], pybind11_stubgen.typing_ext.FixedSize(2) ] ] Kind: Method Returns the skeleton edges as pairs of keypoint indices. ###### getLabelNames(self) -> list [ str ]: list [ str ] Kind: Method Returns the label names for the keypoints. ###### getNumKeypoints(self) -> int|None: int|None Kind: Method Returns the number of keypoints the model detects, or std::nullopt when not configured. ###### getOutputLayerName(self) -> str: str Kind: Method Returns the name of the model output layer to parse. ###### getScaleFactor(self) -> float: float Kind: Method Returns the scale factor to divide the keypoint coordinates by. ###### getScoreThreshold(self) -> float|None: float|None Kind: Method Returns the confidence score threshold for detected keypoints, or std::nullopt when not configured. ###### runOnHost(self) -> bool: bool Kind: Method Returns true when this node runs on the host. Host-only pipelines always run the node on the host. ###### setEdges(self, edges: list [ typing.Annotated [ list [ int ], pybind11_stubgen.typing_ext.FixedSize ( 2 ) ] ]) Kind: Method Sets the skeleton edges as pairs of keypoint indices used for visualizing the skeleton. Example: {{0, 1}, {1, 2}, {2, 3}, {3, 0}} connects keypoint 0 to keypoint 1, keypoint 1 to keypoint 2, etc. Parameter ``edges``: Vector of keypoint index pairs ###### setLabelNames(self, labelNames: list [ str ]) Kind: Method Sets the label names for the keypoints, indexed by keypoint index. Parameter ``labelNames``: Vector of label names ###### setNNArchive(self, nnArchive: depthai.NNArchive) Kind: Method Set NNArchive for this Node. The archive must contain exactly one KeypointParser head; use setNNArchiveHead() to select a specific head from a multi-head archive. Parameter ``nnArchive:``: NNArchive to set ###### setNNArchiveHead(self, head: depthai.nn_archive.v1.Head) Kind: Method Set NNArchive head for this Node. The head must be a KeypointParser head with exactly one output layer. Parameter ``head:``: NNArchive head to set ###### setNumKeypoints(self, nKeypoints: int) Kind: Method Sets the number of keypoints the model detects. Must be configured before the pipeline starts, either explicitly or through an NNArchive head. Parameter ``nKeypoints``: Number of keypoints, must be greater than 0 ###### setOutputLayerName(self, outputLayerName: str) Kind: Method Sets the name of the model output layer to parse. When left empty, the parser selects the tensor automatically if the incoming NNData contains exactly one tensor and fails otherwise. Parameter ``outputLayerName``: Name of the output layer ###### setRunOnHost(self, runOnHost: bool) Kind: Method Select whether the node runs on the host or device. ###### setScaleFactor(self, scaleFactor: float) Kind: Method Sets the scale factor to divide the keypoint coordinates by. Parameter ``scaleFactor``: Scale factor, must be greater than 0 ###### setScoreThreshold(self, threshold: float) Kind: Method Sets the confidence score threshold for detected keypoints. Parameter ``threshold``: Confidence score threshold, must be between 0 and 1 ###### input Kind: Property Input NN results with keypoints data to parse. ###### out Kind: Property Outputs Keypoints message with the parsed 2D or 3D keypoints. ###### depthai.beta.node.LaneDetectionParser(depthai.DeviceNode) Kind: Class LaneDetectionParser node. Parses the output of an Ultra-Fast-Lane-Detection (UFLD) neural network, e.g. the CULane and TuSimple variants, into a dai::beta::Clusters message with one cluster of normalized points per lane, including empty clusters for lanes without enough detected points. The parser consumes a single output tensor. When the incoming NNData contains exactly one tensor, it is selected automatically; otherwise the output layer name must be configured explicitly or through an NNArchive head. The tensor is read in its stored order and must be a 4D tensor of shape (batch, gridingNum + 1, clsNumPerLane, numLanes); the first batch entry is decoded. The row anchors, griding number and number of points per lane must be configured before the pipeline starts, either explicitly or through an NNArchive head. The input size must also be configured before the pipeline starts: building from a full NNArchive derives it from the model input's declared shape and layout (NHWC or NCHW), while building from a specific head requires setInputSize() because a head carries no model input metadata. ###### build() Kind: Method ###### getClsNumPerLane(self) -> int|None: int|None Kind: Method Returns the number of points per lane, or std::nullopt when not configured. ###### getGridingNum(self) -> int|None: int|None Kind: Method Returns the griding number, or std::nullopt when not configured. ###### getInputSize(self) -> tuple [ int, int ]|None: tuple [ int, int ]|None Kind: Method Returns the model input image size as (width, height), or std::nullopt when not configured. ###### getOutputLayerName(self) -> str: str Kind: Method Returns the name of the model output layer to parse. ###### getRowAnchors(self) -> list [ int ]: list [ int ] Kind: Method Returns the row anchors, or an empty vector when not configured. ###### runOnHost(self) -> bool: bool Kind: Method Returns true when this node runs on the host. Host-only pipelines always run the node on the host. ###### setClsNumPerLane(self, clsNumPerLane: int) Kind: Method Sets the number of points per lane. Must be configured before the pipeline starts, either explicitly or through an NNArchive head. Parameter ``clsNumPerLane``: Number of points per lane, must be greater than 0 ###### setGridingNum(self, gridingNum: int) Kind: Method Sets the griding number, the number of column samples the model predicts lane positions over. Must be configured before the pipeline starts, either explicitly or through an NNArchive head. Parameter ``gridingNum``: Griding number, must be greater than 1 ###### setInputSize(self, width: int, height: int) Kind: Method Sets the model input image size the emitted points are computed against and normalized by. Must be configured before the pipeline starts. Configuring from a full NNArchive derives it from the model input's declared shape and layout; the most recent configuration wins. Parameter ``width``: Input image width, must be greater than 0 Parameter ``height``: Input image height, must be greater than 0 ###### setNNArchive(self, nnArchive: depthai.NNArchive) Kind: Method Set NNArchive for this Node. The archive must contain exactly one LaneDetectionParser head and exactly one model input; use setNNArchiveHead() to select a specific head from a multi-head archive. The input size is derived from the model input's declared shape and layout (NHWC or NCHW). Parameter ``nnArchive:``: NNArchive to set ###### setNNArchiveHead(self, head: depthai.nn_archive.v1.Head) Kind: Method Set NNArchive head for this Node. The head must be a LaneDetectionParser head with exactly one output layer. Parameter ``head:``: NNArchive head to set @note A head carries no model input metadata, so the input size must additionally be configured with setInputSize(). ###### setOutputLayerName(self, outputLayerName: str) Kind: Method Sets the name of the model output layer to parse. When left empty, the parser selects the tensor automatically if the incoming NNData contains exactly one tensor and fails otherwise. Parameter ``outputLayerName``: Name of the output layer ###### setRowAnchors(self, rowAnchors: list [ int ]) Kind: Method Sets the row anchors, the image rows at which the model predicts lane positions. Must be configured before the pipeline starts, either explicitly or through an NNArchive head, and must contain at least as many entries as the number of points per lane. Parameter ``rowAnchors``: Row anchors, must not be empty ###### setRunOnHost(self, runOnHost: bool) Kind: Method Select whether the node runs on the host or device. ###### input Kind: Property Input NN results with lane detection data to parse. ###### out Kind: Property Outputs Clusters message with the detected lanes represented as clusters of points. ###### depthai.beta.node.MLSDParser(depthai.DeviceNode) Kind: Class MLSDParser node. Parses the output of the M-LSD line segment detection model into a dai::beta::Lines message with the detected lines and their confidence scores, ordered by descending score. The parser consumes two output tensors that must be configured before the pipeline starts, either explicitly or through an NNArchive head: the tpMap tensor, read in NCHW orientation as a 4D tensor of shape (batch, channels, height, width) whose channels 1 to 4 hold the line displacement maps of the first batch entry, and the heat tensor, flattened to one score per (height, width) grid position. The topK highest-scoring grid positions are decoded into candidate lines and kept when their score and length are strictly above the score and distance thresholds. Ties between equal heat scores are ordered following numpy's portable argpartition/argsort semantics. The emitted line coordinates are normalized by the model input size, which defaults to 512x512 (the input size of all known M-LSD models, hard-coded by the source parser); building from a full NNArchive derives it from the model input's declared shape and layout (NHWC or NCHW), and setInputSize() overrides it. ###### build() Kind: Method ###### getDistanceThreshold(self) -> float: float Kind: Method Returns the distance threshold for detected lines. ###### getInputSize(self) -> tuple [ int, int ]: tuple [ int, int ] Kind: Method Returns the model input image size as (width, height). ###### getOutputLayerHeat(self) -> str: str Kind: Method Returns the name of the output layer containing the heat tensor, or an empty string when not configured. ###### getOutputLayerTPMap(self) -> str: str Kind: Method Returns the name of the output layer containing the tpMap tensor, or an empty string when not configured. ###### getScoreThreshold(self) -> float: float Kind: Method Returns the confidence score threshold for detected lines. ###### getTopK(self) -> int: int Kind: Method Returns the number of top candidates to keep. ###### runOnHost(self) -> bool: bool Kind: Method Returns true when this node runs on the host. Host-only pipelines always run the node on the host. ###### setDistanceThreshold(self, distanceThreshold: float) Kind: Method Sets the distance threshold for detected lines. Candidates whose length in heat map grid units is strictly above the threshold are kept. Parameter ``distanceThreshold``: Distance threshold @note Configures startup behavior. Send MLSDParserConfig to inputConfig after the pipeline starts. ###### setInputSize(self, width: int, height: int) Kind: Method Sets the model input image size the emitted line coordinates are normalized by, x coordinates by the width and y coordinates by the height. Defaults to 512x512, the input size of all known M-LSD models. Configuring from a full NNArchive derives it from the model input's declared shape and layout; the most recent configuration wins. Parameter ``width``: Input image width, must be greater than 0 Parameter ``height``: Input image height, must be greater than 0 ###### setNNArchive(self, nnArchive: depthai.NNArchive) Kind: Method Set NNArchive for this Node. The archive must contain exactly one MLSDParser head and exactly one model input; use setNNArchiveHead() to select a specific head from a multi-head archive. The input size is derived from the model input's declared shape and layout (NHWC or NCHW). Parameter ``nnArchive:``: NNArchive to set ###### setNNArchiveHead(self, head: depthai.nn_archive.v1.Head) Kind: Method Set NNArchive head for this Node. The head must be an MLSDParser head with exactly two output layers; the layer whose name contains "tpMap" is used as the tpMap layer and the layer whose name contains "heat" as the heat layer. Parameter ``head:``: NNArchive head to set @note A head carries no model input metadata, so the input size keeps its current value (512x512 by default); use setInputSize() for models with a different input size. ###### setOutputLayerHeat(self, outputLayerHeat: str) Kind: Method Sets the name of the output layer containing the heat tensor. Must be configured before the pipeline starts, either explicitly or through an NNArchive head. Parameter ``outputLayerHeat``: Name of the output layer containing the heat tensor ###### setOutputLayerTPMap(self, outputLayerTPMap: str) Kind: Method Sets the name of the output layer containing the tpMap tensor. Must be configured before the pipeline starts, either explicitly or through an NNArchive head. Parameter ``outputLayerTPMap``: Name of the output layer containing the tpMap tensor ###### setRunOnHost(self, runOnHost: bool) Kind: Method Select whether the node runs on the host or device. ###### setScoreThreshold(self, scoreThreshold: float) Kind: Method Sets the confidence score threshold for detected lines. Candidates with a heat score strictly above the threshold are kept. Parameter ``scoreThreshold``: Confidence score threshold @note Configures startup behavior. Send MLSDParserConfig to inputConfig after the pipeline starts. ###### setTopK(self, topK: int) Kind: Method Sets the number of top candidates to keep. The number of candidates is capped at the heat map size when decoding. Parameter ``topK``: Number of top candidates to keep, must be positive @note Configures startup behavior. Send MLSDParserConfig to inputConfig after the pipeline starts. ###### initialConfig Kind: Property Configuration used until a message is received on inputConfig. ###### input Kind: Property Input NN results with line detection data to parse. ###### inputConfig Kind: Property Runtime parser configuration. When synchronized, one configuration is consumed per frame; otherwise all queued configurations are drained and the newest valid one is used. ###### out Kind: Property Outputs Lines message with the detected lines and confidence scores. ###### depthai.beta.node.MPPalmDetectionParser(depthai.DeviceNode) Kind: Class MPPalmDetectionParser node. Parses the output of the MediaPipe palm detection model into a dai::ImgDetections message containing the rotated bounding boxes, labels and confidence scores of the detected hands. The decoding is based on https://github.com/geaxgx/depthai_hand_tracker (MIT License). The parser consumes two output tensors and identifies them by their last dimension: the tensor with the larger last dimension holds the raw bounding boxes and is reshaped to (numAnchors, 18) rows of bounding box center/size plus 7 palm keypoint coordinate pairs; the tensor with the smaller last dimension holds the raw scores and is flattened to (numAnchors,). The scores are passed through a sigmoid and filtered with the confidence threshold, the kept rows are decoded against the model's SSD anchors generated from the configured scale (the model input size), converted to rectangles rotated to align the wrist to middle- finger direction with the rectangle's y-axis and expanded to squares, and non- maximum suppression keeps at most the configured maximum number of detections. The emitted bounding boxes are normalized to [0, 1]. ###### build() Kind: Method ###### getConfidenceThreshold(self) -> float: float Kind: Method Returns the confidence score threshold for detected hands. ###### getIouThreshold(self) -> float: float Kind: Method Returns the non-maximum suppression (IoU) threshold. ###### getLabelNames(self) -> list [ str ]: list [ str ] Kind: Method Returns the label names for the detected hands. ###### getMaxDetections(self) -> int: int Kind: Method Returns the maximum number of detections to keep. ###### getOutputLayerNames(self) -> list [ str ]: list [ str ] Kind: Method Returns the names of the model output layers relevant to the parser. ###### getScale(self) -> int: int Kind: Method Returns the scale of the model input image in pixels. ###### runOnHost(self) -> bool: bool Kind: Method Returns true when this node runs on the host. Host-only pipelines always run the node on the host. ###### setConfidenceThreshold(self, threshold: float) Kind: Method Sets the confidence score threshold for detected hands. Detections with a sigmoid score strictly above the threshold are kept. Parameter ``threshold``: Confidence score threshold @note Configures startup behavior. Send MPPalmDetectionParserConfig to inputConfig after the pipeline starts. ###### setIouThreshold(self, threshold: float) Kind: Method Sets the non-maximum suppression (IoU) threshold. Parameter ``threshold``: Non-maximum suppression threshold @note Configures startup behavior. Send MPPalmDetectionParserConfig to inputConfig after the pipeline starts. ###### setLabelNames(self, labelNames: list [ str ]) Kind: Method Sets the label names for the detected hands. The first label name is assigned to every detection (all detections carry label 0). When empty, no label name is assigned. Parameter ``labelNames``: List of label names ###### setMaxDetections(self, maxDetections: int) Kind: Method Sets the maximum number of detections to keep. Parameter ``maxDetections``: Maximum number of detections to keep @note Configures startup behavior. Send MPPalmDetectionParserConfig to inputConfig after the pipeline starts. ###### setNNArchive(self, nnArchive: depthai.NNArchive) Kind: Method Set NNArchive for this Node. The archive must contain exactly one MPPalmDetectionParser head; use setNNArchiveHead() to select a specific head from a multi-head archive. Parameter ``nnArchive:``: NNArchive to set ###### setNNArchiveHead(self, head: depthai.nn_archive.v1.Head) Kind: Method Set NNArchive head for this Node. The head must be an MPPalmDetectionParser head with exactly two output layers. Parameter ``head:``: NNArchive head to set ###### setOutputLayerNames(self, outputLayerNames: list [ str ]) Kind: Method Sets the names of the model output layers relevant to the parser. Exactly two layer names are required. Parameter ``outputLayerNames``: Names of the output layers ###### setRunOnHost(self, runOnHost: bool) Kind: Method Select whether the node runs on the host or device. ###### setScale(self, scale: int) Kind: Method Sets the scale of the model input image in pixels (e.g. 192 for a 192x192 model). The SSD anchors used for decoding are generated from the scale; a scale that does not match the model input size fails decoding with an anchor count mismatch. Parameter ``scale``: Scale of the input image ###### initialConfig Kind: Property Configuration used when the parser starts. ###### input Kind: Property Input NN results with palm detection data to parse. ###### inputConfig Kind: Property Runtime parser configuration. In synchronized mode one configuration is consumed per input frame; otherwise all queued configurations are drained and the newest valid one is retained. ###### out Kind: Property Outputs ImgDetections message with the rotated bounding boxes, labels and confidence scores of the detected hands. ###### depthai.beta.node.MapOutputParser(depthai.DeviceNode) Kind: Class MapOutputParser node. Parses the output of models that produce map outputs, such as depth maps (e.g. DepthAnything), density maps (e.g. DM-Count), heat maps, and similar, into a dai::beta::Map2D message. The parser consumes a single output tensor. When the incoming NNData contains exactly one tensor, it is selected automatically; otherwise the output layer name must be configured explicitly or through an NNArchive head. The tensor is read in its stored order; leading dimensions of 1 are squeezed and the tensor must then be a 2D HW map, or a 3D HWN map with a singleton trailing dimension that is squeezed as well. All map dimensions are derived from the runtime tensor descriptor. When min-max scaling is enabled, the map values are scaled to the [0, 1] range; a constant map is left unchanged. ###### build() Kind: Method ###### getMinMaxScaling(self) -> bool: bool Kind: Method Returns the flag indicating whether the map is scaled to the [0, 1] range. ###### getOutputLayerName(self) -> str: str Kind: Method Returns the name of the model output layer to parse. ###### runOnHost(self) -> bool: bool Kind: Method Returns true when this node runs on the host. Host-only pipelines always run the node on the host. ###### setMinMaxScaling(self, minMaxScaling: bool = True) Kind: Method Sets the flag indicating whether the map is scaled to the [0, 1] range. When true, the map values are min-max scaled to [0, 1]; a constant map is left unchanged. Defaults to false. Parameter ``minMaxScaling``: True to scale the map to the [0, 1] range, defaults to true @note Configures startup behavior. Send MapOutputParserConfig to inputConfig after the pipeline starts. ###### setNNArchive(self, nnArchive: depthai.NNArchive) Kind: Method Set NNArchive for this Node. The archive must contain exactly one MapOutputParser head; use setNNArchiveHead() to select a specific head from a multi-head archive. Parameter ``nnArchive:``: NNArchive to set ###### setNNArchiveHead(self, head: depthai.nn_archive.v1.Head) Kind: Method Set NNArchive head for this Node. The head must be a MapOutputParser head with exactly one output layer. Parameter ``head:``: NNArchive head to set ###### setOutputLayerName(self, outputLayerName: str) Kind: Method Sets the name of the model output layer to parse. When left empty, the parser selects the tensor automatically if the incoming NNData contains exactly one tensor and fails otherwise. Parameter ``outputLayerName``: Name of the output layer ###### setRunOnHost(self, runOnHost: bool) Kind: Method Select whether the node runs on the host or device. ###### initialConfig Kind: Property Configuration used until a message is received on inputConfig. ###### input Kind: Property Input NN results with map tensor data to parse. ###### inputConfig Kind: Property Runtime parser configuration. When synchronized, one configuration is consumed per frame; otherwise all queued configurations are drained and the newest valid one is used. ###### out Kind: Property Outputs Map2D message with the parsed 2D map, e.g. a depth or density map. ###### depthai.beta.node.PPTextDetectionParser(depthai.DeviceNode) Kind: Class PPTextDetectionParser node. Parses the output of the PaddlePaddle OCR text detection model into a dai::ImgDetections message containing the rotated bounding boxes and confidence scores of the detected text. The parser consumes a single probability-map output tensor of shape (1, 1, H, W) or (1, H, W, 1). The map is thresholded with the mask threshold into a binary text mask, the mask is dilated and its contours become rotated-rectangle candidates; when more contours than the maximum number of detections remain, the largest by area are kept. Rectangles smaller than 8 pixels on their smaller side are dropped, each candidate is scored with the mean probability inside its (slightly shrunk) corner polygon, candidates scoring below the confidence threshold are dropped, and the kept rectangles are expanded by sqrt(2) in both dimensions. The emitted bounding boxes are normalized to [0, 1] with angles in degrees rounded to whole numbers; the detections carry no labels. ###### build() Kind: Method ###### getConfidenceThreshold(self) -> float: float Kind: Method Returns the confidence score threshold for the detected text bounding boxes. ###### getMaskThreshold(self) -> float: float Kind: Method Returns the mask threshold for creating the binary text mask from the model output probabilities. ###### getMaxDetections(self) -> int: int Kind: Method Returns the maximum number of candidate bounding boxes. ###### getOutputLayerName(self) -> str: str Kind: Method Returns the name of the model output layer holding the text probability map. ###### runOnHost(self) -> bool: bool Kind: Method Returns true when this node runs on the host. Host-only pipelines always run the node on the host. ###### setConfidenceThreshold(self, threshold: float) Kind: Method Sets the confidence score threshold for the detected text bounding boxes. Candidates with a score strictly below the threshold are dropped. Parameter ``threshold``: Confidence score threshold @note Configures startup behavior. Send PPTextDetectionParserConfig to inputConfig after the pipeline starts. ###### setMaskThreshold(self, maskThreshold: float) Kind: Method Sets the mask threshold for creating the binary text mask from the model output probabilities. Probabilities strictly above the threshold belong to the mask. Parameter ``maskThreshold``: Mask threshold @note Configures startup behavior. Send PPTextDetectionParserConfig to inputConfig after the pipeline starts. ###### setMaxDetections(self, maxDetections: int) Kind: Method Sets the maximum number of candidate bounding boxes. When more candidate contours are found, only the largest by area are kept. Parameter ``maxDetections``: Maximum number of candidate bounding boxes @note Configures startup behavior. Send PPTextDetectionParserConfig to inputConfig after the pipeline starts. ###### setNNArchive(self, nnArchive: depthai.NNArchive) Kind: Method Set NNArchive for this Node. The archive must contain exactly one PPTextDetectionParser head; use setNNArchiveHead() to select a specific head from a multi-head archive. Parameter ``nnArchive:``: NNArchive to set ###### setNNArchiveHead(self, head: depthai.nn_archive.v1.Head) Kind: Method Set NNArchive head for this Node. The head configures the confidence threshold, the mask threshold and the maximum number of detections when present in its metadata. The head's declared output names are not consumed: the parser resolves the single runtime output tensor by itself (or uses the explicitly configured output layer name), mirroring the source parser; archives whose head output name differs from the model's declared output name therefore parse correctly. Parameter ``head:``: NNArchive head to set ###### setOutputLayerName(self, outputLayerName: str) Kind: Method Sets the name of the model output layer holding the text probability map. When empty (the default), the layer is resolved automatically from single-tensor NN results; multi-tensor results require an explicit name. Parameter ``outputLayerName``: Name of the output layer ###### setRunOnHost(self, runOnHost: bool) Kind: Method Select whether the node runs on the host or device. ###### initialConfig Kind: Property Configuration used when the parser starts. ###### input Kind: Property Input NN results with text detection probability map to parse. ###### inputConfig Kind: Property Runtime parser configuration. In synchronized mode one configuration is consumed per input frame; otherwise all queued configurations are drained and the newest valid one is retained. ###### out Kind: Property Outputs ImgDetections message with the rotated bounding boxes and confidence scores of the detected text. ###### depthai.beta.node.RFDETRParser(depthai.DeviceNode) Kind: Class RFDETRParser node. Parses the output of RF-DETR object detection models (https://github.com/roboflow/rf-detr) into a dai::ImgDetections message containing the bounding boxes, labels, confidence scores and, in segmentation mode, an instance segmentation mask, everything normalized to [0, 1]. The parser consumes 2 output tensors for detection (boxes, class logits) or 3 for instance segmentation (boxes, class logits, mask logits), in that order. When no output layer names are configured, all layer names of the incoming NNData are used in their reported order. The boxes tensor squeezes to (N, 4) with normalized (xCenter, yCenter, width, height) boxes, the logits tensor is (1, N, C) and the mask logits tensor squeezes to (N, maskHeight, maskWidth). Class probabilities are the sigmoid of the logits; per query the maximum probability is the score and its class the label. Detections are ordered by descending score, truncated to the maximum number of detections and kept when their score is strictly greater than the confidence threshold. In segmentation mode at most 255 instances fit into the mask, so the truncation is additionally capped at 255. Each detection's mask logits are passed through a sigmoid, cropped to its bounding box, binarized with the mask confidence threshold and resized to the model input size with nearest-neighbor interpolation; the pixels not claimed by an earlier (higher-scoring) detection receive the detection's index, with 255 marking background. ###### build() Kind: Method ###### getConfidenceThreshold(self) -> float: float Kind: Method Returns the confidence score threshold for detected objects. ###### getInputSize(self) -> tuple [ int, int ]|None: tuple [ int, int ]|None Kind: Method Returns the model input image size as (width, height), or std::nullopt when it is not set. ###### getLabelNames(self) -> list [ str ]: list [ str ] Kind: Method Returns the label names for the detected objects. ###### getMaskConfidence(self) -> float: float Kind: Method Returns the mask confidence threshold. ###### getMaxDetections(self) -> int: int Kind: Method Returns the maximum number of detections to keep. ###### getOutputLayerNames(self) -> list [ str ]: list [ str ] Kind: Method Returns the names of the model output layers. ###### runOnHost(self) -> bool: bool Kind: Method Returns true when this node runs on the host. Host-only pipelines always run the node on the host. ###### setConfidenceThreshold(self, threshold: float) Kind: Method Sets the confidence score threshold for detected objects. Detections with a score strictly greater than the threshold are kept. Defaults to 0.5. Parameter ``threshold``: Confidence score threshold, must be between 0 and 1 @note Configures startup behavior. Send RFDETRParserConfig to inputConfig after the pipeline starts. ###### setInputSize(self, width: int, height: int) Kind: Method Sets the model input image size the segmentation mask is emitted at. Unset by default; segmentation mode requires it to be configured from an NNArchive or with this setter before the parser processes messages. Detection mode does not use it. Configuring from a full NNArchive derives it from the first model input's declared shape and layout; the most recent configuration wins. Parameter ``width``: Input image width, must be greater than 0 Parameter ``height``: Input image height, must be greater than 0 ###### setLabelNames(self, labelNames: list [ str ]) Kind: Method Sets the label names for the detected objects, indexed by the class label. A detection whose label is out of range receives the name "class_