vernier-core

Pure-Rust core for the vernier evaluation library.

This crate is the source of truth for vernier's instance-paradigm evaluation semantics — bounding-box AP, segmentation AP, boundary-IoU AP, and OKS keypoints. It links no Python and is usable directly from Rust binaries, CLI tools, or embedded contexts.

It also covers two cross-paradigm surfaces:

Streaming — fold per-image evals incrementally and finalize on demand (ADR-0013).
Distributed partials — encode/decode the rkyv wire format for sharding evaluation across ranks (ADR-0031, ADR-0032).

Why a separate crate

The Python wheel (pip install vernier) is a thin PyO3 shim over this crate. By design, every evaluation decision lives here. vernier-ffi is a data-conversion layer; vernier-cli orchestrates eval invocations; both depend on vernier-core and never the other way around.

If your project is pure-Rust — a robotics replay pipeline, a Rust-side training benchmark, a custom evaluation tool — vernier-core is the entry point.

Installation

[dependencies]
vernier-core = "0.0.1"

Stable Rust only (MSRV in rust-toolchain.toml, currently 1.83).

Minimal usage

use vernier_core::{
    evaluate_bbox, AreaRange, CocoDataset, CocoDetections, EvalDataset,
    EvaluateParams, ParityMode,
};

let gt: CocoDataset = serde_json::from_slice(&std::fs::read("gt.json")?)?;
let dt: CocoDetections = serde_json::from_slice(&std::fs::read("dt.json")?)?;
let params = EvaluateParams::default()
    .with_parity_mode(ParityMode::Strict)
    .with_max_dets(vec![1, 10, 100]);

let summary = evaluate_bbox(&gt.into_eval_dataset(), &dt, &params)?;
println!("{}", summary.pretty_lines().join("\n"));

A worked example is in examples/cache_speedup_val2017.rs.

Parity contract

vernier-core reproduces pycocotools==2.0.11 semantics under the parity model in ADR-0002. Each documented quirk is dispositioned as strict (bit-equal output) or corrected (opt-in opinionated fix). The full disposition table lives in docs/engineering/pycocotools-quirks.md.

License