vernier-cli

Command-line driver for the vernier evaluation library.

The vernier binary runs COCO-format evaluation (bbox, segm, boundary, keypoints) without a Python interpreter. CI pipelines, robotics replay pipelines, and shell-driven non-regression checks are the audience — ADR-0015 captures the design.

Installation

Pre-built binaries (recommended for CI):

# cargo binstall — downloads the right binary for your host triple
cargo binstall vernier-cli

# Or curl-pipe-bash from the GitHub Release
curl -L https://github.com/NoeFontana/vernier/releases/latest/download/vernier-installer.sh | bash

From source:

cargo install vernier-cli

Pre-built binaries ship for x86_64-unknown-linux-gnu, aarch64-unknown-linux-gnu, aarch64-apple-darwin, and x86_64-pc-windows-msvc.

Usage

# Strict-mode parity with pycocotools' summarize() stdout
vernier eval --gt gt.json --dt dt.json --iou-type bbox

# Multi-emit: one eval, multiple outputs
vernier eval --gt gt.json --dt dt.json --iou-type segm \
    --emit text \
    --emit json=result.json

Output is byte-deterministic across runs of the same input — no timestamps, no host metadata, sorted JSON keys, atomic file writes. CI archives diff cleanly and regression-tracking buckets work out of the box.

The full surface (--parity-mode, --max-dets, --use-cats, --dilation-ratio, --sigmas, --emit FMT[=PATH], --quiet) is documented in ADR-0015 §"Surface".

Exit codes

Code	Meaning
0	Eval ran, summary written
1	Eval ran but failed (parity-mode constraint, malformed JSON, kernel error)
2	Argument parse / validation failure

Architectural notes

vernier-cli is a thin parse-and-format layer over vernier-core. It links neither vernier-ffi nor PyO3 — pure-Rust dep tree per ADR-0015. The binary's stability commitments (flag surface, JSON schema version, output determinism) are pinned from its first ship under the 0.0.x release line.

License