Components

Track Objects (Hybrid SORT)

@pipelogic

Hybrid-SORT multi-object tracker over per-frame [BoundingBox] or [Segmentation]. It extends observation-centric SORT with weak cues - a confidence-state score and a velocity-direction term - so heavily overlapping boxes in a crowd are still told apart. A Kalman motion model plus class-aware gating give stable ids. Emits one active-track list per frame, not per box.

Detect Objects by Visual Prompt (Ultralytics YOLOE)

@ultralytics

Embeds example regions from the bound reference at refer_boxes into a YOLOE visual prompt, then finds matching objects in each target Image. The consumer picks one arm: boxes, rectangles, classes, masks with class, bare masks, count, or annotated frame.

Estimate Pose (Ultralytics YOLO)

@ultralytics

Human-pose estimator (Ultralytics YOLO-pose) over each Image, emitting a deploy-time-resolved shape: flat [Landmark] keypoints, per-person [[Landmark]], per-person [([Landmark], BoundingBox)], or [Skeleton] with COCO-17 bone topology.

Write Database (SQLite)

@sqlite

SQLite row writer. Accepts any record { $ns: $ts... } and INSERTs one row per tick into a configured table, auto-creating schema from field names when enabled. Non-atomic values serialise as JSON TEXT or drop the row. Pair with a SQLite reader to serve the rows.

Segment Image (Mask2Former)

@meta

Universal image segmentation on an Image with a Mask2Former checkpoint. The task setting selects the panoptic, instance or semantic head; the sink resolves one of nine projections of the per-segment result, from [(Segmentation, BoundingBox)] to a count.

Extract Text (Tesseract OCR)

@tesseract

Runs the classic CPU Tesseract engine over each Image, emitting either a joined String transcript or per-word regions with Rectangle bounds and optional confidence normalised to 0..1.

Recognize Action Skeleton (STGCN PP)

@kennymckormick

Buffers a sliding window of per-frame Landmarks2d.Human22 skeletons and runs an STGCN++ graph-convolution network over the raw joint sequence to emit a ranked [Classification] or a single top-1 Classification of NTU RGB+D 60 actions.

Edit Image (HuggingFace)

@huggingface

Edits each input Image under a text prompt with a HuggingFace diffusers image-to-image checkpoint (default stabilityai/sdxl-turbo) and emits one edited Image at the same resolution. strength sets how far the edit departs from the source.

Output Parquet File

Writes each upstream Analytics.TabularDataset to a columnar Parquet file under output_path, naming it from name_template and applying the column codec named in compression.

Remove Background (BiRefNet)

@zhengpeng

Salient-subject background removal on an Image with BiRefNet. A soft matte is predicted and the sink resolves the arm: an RGBA cutout Image with the soft alpha, a (Segmentation, BoundingBox), a (Mask, DetectedClass), or a [Mask].

Analyze Image (Ollama VLM)

@ollama

Sends one or more images plus a String prompt to a vision-language model served by a local Ollama sidecar and returns the model's reply as a String. The image input is a oneof spanning a single image, an image-with-label, a batch, or labeled batch.

Threshold Image (OpenCV)

@opencv

Binary thresholding over Image using fixed, Otsu, adaptive-mean, adaptive-Gaussian, or triangle methods. Converts input to greyscale and emits either a single-channel Image foreground mask or, when consumed by a mask-typed downstream, a boolean Mask.

Extract Edges (Canny)

@opencv

Canny edge extractor: converts each Image to greyscale, runs hysteresis edge detection over a configurable threshold pair and Sobel aperture, and emits a binary edge Mask aligned to the input frame.

Detect Objects (RF DETR)

@roboflow

Runs a bound RF-DETR checkpoint locally over each Image, building one per-detection record and projecting it onto the connected output shape: boxes, geometry, labels, a count, or an overlay, filtered by confidence.

Triangulate 3D Points (Depth)

@bleenco

Monocular depth unprojection (NOT multi-view triangulation). Bilinearly samples a MetricDepthMap at each [Point] and inverse-pinhole back-projects through a CameraIntrinsics into camera-frame 3D.

Measure Audio Quality (DNSMOS)

@microsoft

Scores each AudioFrame with the DNSMOS network and emits four MOS-style quality numbers (speech, background, overall, and a P.808 estimate), shaped either as a named record or a [Analytics.Metric] list depending on what the downstream stage expects.

Input Image HTTP

Exposes a public image ingress that accepts client-pushed bytes over HTTP POST or WebSocket, decodes them, and emits an Image paired with a per-upload session-id String. Use it when a deployment needs to take images from outside the graph.

Apply Image Color Correction (OpenCV)

@opencv

Per-pixel tonal and color adjustment on an Image, returning a corrected Image: applies contrast and brightness, then a gamma tone curve, then HSV saturation and hue_shift rotation, in that fixed order.