Crate warden_vision 

warden-vision: reusable computer vision building blocks for detection, tracking, and annotation, mirroring the API shape of Roboflow’s Python supervision library.

Modules

• core: core::Detection / core::Detections, the shared currency between detectors, trackers, zones, and annotators, plus core::KeyPoints (pose/landmark data, core::COCO_17_EDGES skeleton), core::Classifications (whole-image classification), core::InferenceSlicer (tiled/SAHI-style inference), and core::Detections::from_ultralytics_onnx / core::Detections::from_transformers_onnx for parsing raw detector output tensors (see their docs for why these parse tensors rather than binding to ultralytics/transformers objects).
• geometry: geometry::Point, geometry::Rect, the anchor-point geometry::Position enum, polygon utilities (geometry::polygon_area, geometry::polygon_centroid, geometry::polygon_to_rect), the geometry::Zone trait, and its geometry::LineZone / geometry::PolygonZone implementations.
• tracker: tracker::SortTracker (single-stage Kalman + IoU) and tracker::ByteTracker (two-stage high/low-confidence matching, after Zhang et al.’s ByteTrack), both assigning core::Detection::tracker_id; tracker::DetectionsSmoother to damp jitter in a tracked box across frames.
• annotators: the annotators::Annotator trait and a family of drawing configs — boxes (round-cornered, corner-only), masks, polygons, labels (bitmap font, or under annotate-image real TTF/OTF via annotators::RichLabelAnnotator), circles, dots, triangles, ellipses, motion traces, heat maps, halos, percentage bars, icon overlays, background dimming, blur/pixelate redaction, zone outlines, and keypoint skeletons — with pure-Rust (annotate-image) and OpenCV (annotate-opencv) rendering backends (RichLabelAnnotator is pure-Rust only; see its docs for why).
• metrics: metrics::Precision, metrics::Recall, metrics::F1Score, metrics::ConfusionMatrix, metrics::MeanAveragePrecision, and metrics::MeanAverageRecall for scoring predicted core::Detections against ground truth.
• utils: utils::FPSMonitor for throughput reporting, plus (under annotate-image) utils::resize_keeping_aspect_ratio, utils::letterbox, utils::crop_image, utils::overlay_image, and utils::ImageSink for writing a numbered frame sequence.
• dataset (under datasets): dataset::DetectionDataset plus COCO (dataset::coco), YOLO (dataset::yolo), and Pascal VOC (dataset::pascal_voc) loaders/exporters, and dataset::ClassificationDataset with an image-folder-per-class loader/exporter (dataset::classification).
• video (under annotate-opencv): video::VideoInfo, video::frames, and video::VideoSink for reading and writing video files via OpenCV’s videoio.

(Items behind a non-default feature are written above as plain code text rather than intra-doc links, since those links only resolve when this crate’s docs are built with that feature enabled. The published docs build with annotate-image and datasets — see [package.metadata.docs.rs] in Cargo.toml — but not annotate-opencv, since docs.rs has no system OpenCV install; read video’s module docs directly in src/video/mod.rs instead.)

Scope

This crate targets the core, dependency-light building blocks of Roboflow’s supervision: detection/keypoint/classification data structures, geometry, tracking, annotation, evaluation metrics, dataset I/O, image utilities, and OpenCV-backed video I/O. Remaining known gaps: RichLabelAnnotator has no annotate-opencv implementation (OpenCV has no built-in TTF rendering without FreeType); a dedicated OrientedBoxAnnotator doesn’t exist since annotators::PolygonAnnotator already draws an arbitrary 4-point rotated box stored in core::Detection::mask; and detector-output parsing covers Ultralytics YOLO and DETR-family transformers models (the two most common) rather than every framework supervision has an adapter for.

Re-exports

pub use error::Error;

pub use error::Result;

Modules

annotators

Backend-agnostic annotation: drawing detections and zone counters onto an image or video frame.

core

Core data structures: detections and the value types built on top of them.

dataset

Dataset loaders and exporters: three object-detection annotation formats (coco, yolo, pascal_voc) built around the shared DetectionDataset shape, plus a folder-per-class classification layout (classification) built around ClassificationDataset.

error

Shared error type for warden-vision.

geometry

Spatial primitives and zones: points, counting lines, and polygonal regions of interest.

metrics

Evaluation metrics for comparing predicted crate::core::Detections against ground-truth annotations: Precision, Recall, F1Score, ConfusionMatrix, MeanAveragePrecision, and MeanAverageRecall.

tracker

Multi-object tracking: Kalman motion prediction and IoU-based association.

utils

Small standalone utilities that don’t belong to any single pipeline stage: FPSMonitor for throughput reporting, and (under the annotate-image feature) pure-Rust image helpers.