Components

Detect Landmarks (RF DETR)

@roboflow

Detects every person in a single Image and emits their COCO-17 body keypoints in one RF-DETR forward pass — box and keypoints predicted jointly, no separate detector or top-down crop stage — packaged per person as [[Landmark]].

Compare Embedding Pairs

@bleenco

Computes the cosine similarity between two Embedding vectors and emits either the raw Double score or a Bool match decision once the score clears the configurable threshold.

Denoise Audio (FRCRN)

@modelscope

FRCRN speech denoiser (ClearerVoice-Studio): mixes each in_0 AudioFrame to mono, resamples to 16 kHz, and emits a denoised 16 kHz mono AudioFrame on out_0. Use ahead of narrowband ASR or VAD.

Segment Image (Triton)

@nvidia

Image instance segmentation on an Image, served by a Triton inference service; the connected sink resolves the output arm (masks, boxes, parallel arrays, a merged mask, a count, or a colorized image).

Unwarp Document (docTR)

@haofeng

Rectifies a heavily curled or perspective-distorted document Image with a segmentation mask plus the GeoTr transformer, which predicts a backward map that is upsampled and applied to emit a flattened Image for OCR.

Log Message

Terminal debug sink that prints every incoming message of any stream t to standard output, prefixed with a per-process sequence index, so a backend stage's output can be inspected during development. Consumes the stream; nothing flows past it.

Output XLSX File

Writes each upstream Analytics.TabularDataset to an .xlsx workbook under output_path, placing the rows on a worksheet named from sheet_name and naming the file from name_template.

Unwarp Document (DocScanner)

@haofeng

Rectifies a curved or perspective-distorted document Image with the DocScanner-L pipeline: a salient-region segmenter masks the page, an iterative rectifier predicts a backward map, and the warped Image is emitted for downstream OCR.

Classify Image (Ultralytics YOLO)

@ultralytics

Classifies an Image with a local Ultralytics YOLO checkpoint; the ranked result projects onto the connected arm: Classification, [Classification], String, [DetectedClass], UInt64, or [Double].

Smooth 2D Positions (Kalman Filter)

@bleenco

Per-identity constant-velocity Kalman smoother. It takes [(UInt64, Point)] id-and-position pairs and emits a smoothed position with an averaged velocity direction per id as [Vector2d], holding one filter per tracking id across frames.

Output YAML File

Parses each upstream String (or every element of a [String]) as JSON and writes the result as a YAML document under output_path, with the extension from file_extension and the file named from name_template.

Recognize Hand Gesture

@google

Names the hand gesture in an Image with the MediaPipe gesture recognizer, emitting one labelled confidence per detected hand as a ranked [Classification], or the single leading Classification, whichever the downstream connection requires.

Extract Tables TATR

@microsoft

Detects table structure in a document Image with Microsoft Table Transformer and emits either [Rectangle] geometry boxes or [BoundingBox] with class ids, filtered by score_threshold. Use for table layout extraction from pages.

Convert Depth to Point Cloud

@bleenco

Back-projects a MetricDepthMap through a pinhole camera into 3D, emitting a dense per-pixel PointMap3D or a filtered PointCloud. Uses a wired CameraIntrinsics or a focal-length fallback, and can colour the cloud from an optional Image.

Detect Image Anomaly (PatchCore)

@amazon

PatchCore anomaly detector: takes an Image and emits {score: Double, heatmap: Image} — an image-level anomaly score plus a per-pixel heatmap Image. Sharper localisation than PaDiM at higher cost from coreset memory-bank matching over backbone features.

Locate Bodypart

Crops named body regions out of an Image guided by one or more [Landmarks2d.Human22] poses, emitting a per-person [(Image, Double)] of region image plus a confidence so a downstream classifier can work on the isolated part rather than the whole frame.

Translate Text (Opus MT)

@helsinki_nlp

Translates a String or a [String] batch with a Helsinki-NLP Opus-MT (MarianMT) checkpoint and emits text, a list, or a Translation per the resolved output. The language pair is fixed by the bound checkpoint; light enough for CPU.

Dereverberate Audio (NARA WPE)

@fgnt

Deterministic WPE dereverberation (no neural network, no GPU): each AudioFrame is mixed to mono, resampled to 16 kHz, dereverberated in the STFT domain, and emitted as a 16 kHz AudioFrame. Removes late room reflections only — it does not remove noise.