Namespace VisioForge.Core.Types.X.AI

Classes

BackgroundRemovalSettings

Settings for the background-removal (matting) block/engine.

ByteTrackerSettings

Settings for the ByteTrack multi-object tracker.

DetectionFilterSettings

Settings for filtering detections before they are passed to the tracker.

FaceGallery

An in-memory gallery of enrolled face identities used for 1:N face recognition. Each identity holds one or more L2-normalized face embeddings; a query embedding is matched by cosine similarity (the maximum over all stored embeddings of all identities). The gallery is thread-safe and can be persisted to and loaded from disk.

FaceRecognitionSettings

Settings for the face recognition (face identity) block. The block runs a two-stage pipeline: a face detector (YuNet) locates faces and their five landmarks, each face is aligned to a canonical 112x112 crop and turned into an embedding (SFace / ArcFace), and the embedding is matched 1:N against an enrolled VisioForge.Core.Types.X.AI.FaceGallery by cosine similarity.

FrameEmbeddingIndex

An in-memory index of per-frame CLIP embeddings used for semantic video search. Each entry stores a frame's presentation timestamp, the source tag it came from, and the embedding vector produced by a VideoEmbeddingBlock. A query embedding (for example the CLIP text embedding of a natural-language phrase) is matched by cosine similarity and the closest frames are returned. The index is thread-safe and can be persisted to and loaded from disk.

FrameEmbeddingSearchResult

A single hit returned by VisioForge.Core.Types.X.AI.FrameEmbeddingIndex.Search(System.Single[],System.Int32): the video timestamp of an indexed frame, the source tag it was indexed under, and the cosine similarity score against the query embedding.

LicensePlateRecognizerSettings

Settings for the license plate recognizer (ANPR/LPR) block. The block runs a specialized two-stage pipeline: a dedicated license-plate detector (YOLO) locates plates in the frame, then a plate-specific OCR model reads the characters of each cropped plate.

LineZoneSettings

Settings for a directed line (tripwire) zone.

ObjectAnalyticsOverlaySettings

Settings controlling the visual overlay rendered by object analytics.

ObjectAnalyticsSettings

Aggregate settings for the object analytics block, combining detector, tracker, filter, overlay, line-zone, and polygon-zone configuration.

OcrSettings

Settings for the OCR block, which runs a multi-stage PaddleOCR (PP-OCR) pipeline on video frames: text detection (DBNet) → optional angle classification → text-line recognition (CRNN/SVTR + CTC).

OnnxInferenceSettings

Settings for a generic ONNX inference block/engine.

OpenVocabularyDetectorSettings

Settings for the open-vocabulary object detector block/engine (OWLv2 and Grounding DINO).

PIIFaceRedactionSettings

Face-category settings for the PII redaction block. Faces are located with the OpenCV Zoo YuNet detector (MIT); only detection runs — no identity recognition.

PIIPlateRedactionSettings

License-plate-category settings for the PII redaction block. Plates are located with the FastALPR YOLOv9-T end-to-end detector (MIT); only detection runs — the plate characters are never read.

PIIRedactionSettings

Settings for the PII redaction block, which automatically obscures personally identifiable information in video frames: faces (YuNet), vehicle license plates (FastALPR detector), and on-screen text (PaddleOCR detector, with an optional regex filter on the recognized text). Each category can be enabled independently and all use detection only — nothing is identified, recognized, or exported unless the text regex filter explicitly requires recognition.

PIITextRedactionSettings

On-screen-text-category settings for the PII redaction block. Text regions are located with the PaddleOCR PP-OCR detection model and confirmed with the recognition model (both Apache-2.0): recognition is what filters the detector's false positives (on a real scene the detector alone fires on textures and edges, so detection without recognition would redact most of the frame). By default every recognized text region is redacted; set VisioForge.Core.Types.X.AI.PIITextRedactionSettings.TextFilterRegex to redact only the regions whose recognized text matches (for example e-mail addresses or phone numbers). The detection, recognition, and dictionary models are all required whenever VisioForge.Core.Types.X.AI.PIITextRedactionSettings.Enabled is true.

PolygonZoneSettings

Settings for a polygon zone used for occupancy counting.

SileroVadSettings

Settings for the Silero neural voice-activity detector that segments speech inside the speech-to-text block.

SpeechSegment

A single recognized speech segment: a span of transcribed text with its start/end time on the media timeline.

SpeechToTextSettings

Settings for the speech-to-text block (Whisper ASR + optional Silero VAD).

VLMRegion

A single grounded region produced by a Florence-2 task: a text label (an object category, a region description, or a recognized text block) together with its bounding box in source-frame pixel coordinates.

VLMSettings

Settings for the Florence-2 vision-language model (VLM) block. The block runs a four-session ONNX pipeline (vision encoder, token embedder, text encoder, and a merged decoder) to caption frames, detect objects, run OCR, or ground phrases, selected by VisioForge.Core.Types.X.AI.VLMSettings.Task.

VideoEmbeddingSettings

Settings for the video embedding block, which encodes sampled video frames into CLIP image embeddings for semantic search. The block runs a CLIP dual-tower model: a vision tower turns each sampled frame into an L2-normalized embedding, and a text tower (used via EncodeText) turns a natural-language query into an embedding in the same space so the two can be compared by cosine similarity.

YoloDetectorSettings

Settings for the YOLO object detector block/engine.

Enums

BackgroundRemovalModel

The segmentation/matting model family the background-removal block runs. Each family expects a distinct input size and pixel normalization, so the block must be told which one the supplied .onnx file belongs to. The model produces a single-channel foreground alpha mask that the block uses to composite the replacement background.

BackgroundReplacementMode

How the background-removal block replaces the pixels the matting model marks as background (low foreground alpha).

DetectionAnchor

Defines which point of a detection bounding box is used as its reference anchor for line-crossing and polygon-zone evaluation.

FaceEmbeddingModel

The face-embedding model family used by the face recognizer. The family only selects the input preprocessing of the aligned 112x112 crop (channel order and normalization); the embedding length is always read from the model itself, so any vector size works.

LineCrossingDirection

The direction of a directed line crossing.

ObjectDetectorModel

The object-detection model family the detector decodes. Each family has a distinct ONNX output layout and a distinct frame-preprocessing convention (resize mode, pixel normalization, channel order), so the detector must be told which one the supplied .onnx file belongs to.

OnnxExecutionProvider

Execution provider used by the ONNX Runtime inference session.

OpenVocabularyModel

The open-vocabulary detection model family a supplied ONNX model belongs to. This selects the text tokenizer, the frame-preprocessing convention, and the output decoder used by the open-vocabulary detector engine.

PIICategory

The category of personally identifiable information a redacted region belongs to.

PIIRedactionStyle

The visual style used to redact a detected PII region in the video frame.

SpeechToTextTask

The Whisper task: plain transcription or speech translation to English.

VLMTask

The task a Florence-2 vision-language model performs on each processed frame. The task selects the natural-language prompt fed to the model and how its output is interpreted (free text vs. grounded regions).

WhisperModelSize

The Whisper (GGML) model variant. Larger models are more accurate but slower and use more memory; quantized variants (Q5/Q8) trade a little accuracy for a smaller footprint and faster CPU inference.

ZoneExitReason

The reason a tracked object exited a polygon zone.