Crate hotcoco 

Perception evaluation in pure Rust.

Detection ships today — bbox, segmentation, keypoints, and oriented boxes across the COCO, LVIS, and Open Images protocols — on a layered engine that other metric families will share.

use hotcoco::{COCO, COCOeval, params::IouType};
let gt = COCO::new(std::path::Path::new("instances_val2017.json"))?;
let dt = gt.load_res(std::path::Path::new("detections.json"))?;

let mut ev = COCOeval::new(gt, dt, IouType::Bbox);
ev.run();                       // evaluate -> accumulate -> summarize
let report = ev.report()?;      // metrics, per-class, curves, provenance

How the crate is laid out

Module	What lives there
types	The COCO schema — Dataset, Image, Annotation, Category, Rle.
coco	The dataset object: load, index, query, filter, merge, split, sample.
mask, geometry	RLE codec and rotated-rect mechanics.
primitives	Matching kernels — similarity, greedy assignment, LSAP.
metrics	Metric functions over flat arrays — AP, calibration, confusion, bootstrap.
report	EvalReport — the shape every metric family reports in.
detection	The detection metric family: AP/AR, LVIS, Open Images, TIDE.
quality	Dataset introspection: health checks and statistics.
convert	YOLO, Pascal VOC, CVAT, DOTA, and Open Images conversion.

The functional layer

primitives and metrics are free functions over flat arrays — no evaluator required, the way sklearn.metrics and torchmetrics.functional work:

use hotcoco::metrics::counts::average_precision;

let ap = average_precision(&[0.9, 0.8, 0.3], &[true, false, true], None, 3, &[0.0, 0.5, 1.0]);

The two split by what a function produces: primitives produces matches, metrics produces numbers from matches (see metrics for the full split). Between them they hold the implementation of every similarity, matching and accumulation rule in the crate — exactly one of each, with tests/architecture.rs failing the build if a second appears. That is what makes them the place for an auditor to look.

COCOeval is the stateful driver on top — it owns the pycocotools-compatible evaluate/accumulate/summarize lifecycle, and its analysis methods are adapters that marshal eval_imgs into arrays and call the functions above.

Coming from 0.x

1.0 renamed several module paths (eval → detection and friends) with no aliases; the crate-root re-exports resolve unchanged. The rename table is in the migration guide.

Re-exports

pub use coco::COCO;

pub use convert::ConvertError;

pub use convert::CvatImportStats;

pub use convert::CvatStats;

pub use convert::DotaStats;

pub use convert::OidStats;

pub use convert::VocStats;

pub use convert::YoloStats;

pub use detection::AccumulatedEval;

pub use detection::AnnotationIndex;

pub use detection::COCOeval;

pub use detection::CalibrationResult;

pub use detection::CategoryDelta;

pub use detection::CompareOpts;

pub use detection::ComparisonResult;

pub use detection::ConfusionMatrix;

pub use detection::DtStatus;

pub use detection::ErrorProfile;

pub use detection::EvalImg;

pub use detection::EvalMode;

pub use detection::EvalParams;

pub use detection::EvalResults;

pub use detection::EvalShape;

pub use detection::FreqGroup;

pub use detection::GtStatus;

pub use detection::ImageDiagnostics;

pub use detection::ImageSummary;

pub use detection::LabelError;

pub use detection::LabelErrorType;

pub use detection::MetricDef;

pub use detection::SliceResult;

pub use detection::SlicedResults;

pub use detection::TideErrors;

pub use detection::compare;

pub use error::Error;

pub use primitives::greedy::ThreshMatrix;

pub use detection::hierarchy::Hierarchy;

pub use metrics::bootstrap::BootstrapCI;

pub use metrics::calibration::CalibrationBin;

pub use params::AreaRange;

pub use params::IouType;

pub use params::Params;

pub use quality::CategoryStats;

pub use quality::DatasetStats;

pub use quality::DatasetSummary;

pub use quality::Finding;

pub use quality::HealthReport;

pub use quality::Layer;

pub use quality::SummaryStats;

pub use report::EvalReport;

pub use report::Provenance;

pub use types::Annotation;

pub use types::Category;

pub use types::Dataset;

pub use types::Image;

pub use types::Rle;

pub use types::Segmentation;

Modules

coco

COCO dataset loading and querying API.

convert

Format converters: COCO ↔ YOLO, Pascal VOC, CVAT, DOTA, and Open Images.

detection

The detection metric family.

error

geometry

Computational geometry primitives for oriented bounding box (OBB) evaluation.

mask

Pure Rust implementation of COCO mask operations: RLE encoding and decoding, IoU, merge, and area.

metrics

Metric functions — numbers computed from matches.

params

primitives

Matching kernels — the shared substrate that decides what pairs with what.

quality

Dataset quality and introspection — a tier of its own.

report

EvalReport — the shape every family’s results take.

types