Command line reference

Commands are flat: st <command> [flags], with st container <action> and st docs bundle the two subgroups. Arguments after a bare -- are forwarded to the underlying binary.

Configuration is layered: built-in defaults → cli/config.toml → cli/config.local.toml → ST_* environment variables (ST_BUILD_DIR, ST_BUILD_TYPE, ST_GENERATOR, ST_CMAKE, ST_DOCKER_IMAGE, ST_TRACKER, …). st config prints which layer each value came from.

st build

st test

Runs tests through CTest labels. Each test target carries one label, assigned in tests/*/CMakeLists.txt on its gtest_discover_tests call, so a new test file, or a whole target, is picked up with no CLI change.

st test                              # functional: unit + integration + regression
st test --suite unit
st test --suite unit -f 'RectTest'
st test --suite integration --repeat 3
st test --suite all -j 8

For GTest-level knobs (shuffle, break-on-failure) run the binary directly instead:

st run unit_test -- --gtest_shuffle --gtest_break_on_failure

st eval

The MOT label benchmark. Not a CTest suite: it runs regression_test over det.txt-backed scenarios, parses the per-sequence [ PERF ] timings out of its output, then runs TrackEval and prints a metrics table (HOTA / DetA / AssA / MOTA / IDF1 / IDSW / FP / FN / Lat). st test --suite regression answers only pass/fail for the same binary.

st eval
st eval --compare
st eval --scenario MOT17-04-DPM --compare
st eval --eval-only --export ./results
st eval --trackeval --private

The internal runner needs the evaluator’s own dependencies (numpy, scipy, TrackEval) from environment.yml; --trackeval bypasses it. Scenarios registered in tests/regression/ScenarioList.h: full MOT17 train set (21 sequences across DPM/FRCNN/SDP detectors plus -YOLOX variants for each), MOT20 (-01/-02/-03/-05 with -YOLOX variants), and DanceTrack sequences (dancetrack0001, 0002, 0006, 0008, 0012, 0015, …).

st infer

Launches the YOLOX-driven video inference pipeline (tests/regression/inference/demo.py), which loads tracker_bridge.{dll,so} via ctypes. Arguments after -- are forwarded to demo.py untouched.

st infer -n yolox-x -c third_party/weight/mot17x.pth.tar --source third_party/video/MOT17-04.mp4 --num-classes 1 --legacy

st infer -n yolox-x -c third_party/weight/mot17x.pth.tar --source tests/regression/evaluator/data/MOT20-01/img1 --num-classes 1 --legacy

st infer --source 0 --tracker sushitrack --no-save --no-screen --stress

--source also accepts an image-sequence directory (<path>/img1), not just a video file or camera index. tests/regression/evaluator/data/ already holds 76 real tracker benchmark sequences (MOT17, MOT20, DanceTrack) used by the regression suite. Any of their names works, e.g. tests/regression/evaluator/data/MOT20-01/img1 or tests/regression/evaluator/data/dancetrack0001/img1.

If the YOLOX checkpoint filename contains mot17 and --tsize is omitted/640, the pipeline overrides tsize to (800, 1440) to match the MOT17 training resolution.

st clean

Removes the build directory and the assembled package directory (build/ and package/ by default; both overridable in the paths section of cli/config.toml).

st clean

st run

Runs an executable from the build tree. The target is matched by exact name first, then by substring; with no target, or with --select, st lists the executables and asks which one to run. Arguments after -- go to the executable.

st run
st run unit_test
st run unit_test -- --gtest_filter='TrackletTest.*'

st deploy

st deploy {cpp,python} assembles package/ from the current build without building first; st build --deploy {cpp,python} does both. The layouts are under Deploy packages.

st deploy cpp
st deploy python

st container build / st container run

docker build / docker run are issued directly; there are no shell wrappers. run mounts the repo at the configured docker.workdir and passes --gpus all unless --no-gpu is given: docker run -it --rm --gpus all -v "<root>:/workspace/sushitrack" -w /workspace/sushitrack sushitrack bash. Args after -- replace the default bash with the command run inside the container.

st container build
st container build --no-cache
st container run
st container run -- bash -lc "echo hi"

st docker build and st docker run are the old spellings. They still run, are left out of st --help and st --describe, print one notice naming the new spelling when standard error is a terminal, and are removed after 2027-03-25. ST_NO_DEPRECATION_NOTICE=1 silences the notice.

st docs bundle

Packs the manual pages docs/publish.toml lists into docs-bundle-<release>.tar.gz, the archive docs.sushisystems.io reads, and writes docs-bundle-<release>.tar.gz.sha256 beside it. The command comes from sushicore and behaves the same in every Sushi CLI.

It prints the archive’s path with its page count, then sha256 and the digest. The bundle records the checked-out commit, so the command needs a git checkout. SushiTrack publishes no API reference: the bundle holds pages alone, and a bare st docs prints the group’s help and exits with code 2.

Every Markdown file under getting_started/, guides/, architecture/ and reference/ is published. The command stops with one line and exit code 1 when a page has no level-one heading, when docs/README.md does not link it, or when one of its links names a file that does not exist.

st docs bundle --release 1.2.3
st docs bundle --release 1.2.3 --out build/site

Diagnostics

All four are read-only.

Provisioning

setup, doctor, link and unlink come from sushicore and behave the same in every Sushi CLI.

Root options

Exit codes