v0.1.1

Replaces v0.1.0, whose package could not complete its first install.

What was wrong with v0.1.0

Two defects, found one after the other by installing the published package on a clean Ubuntu 24.04 with
its own Docker daemon and running the install end to end.

The image build failed. The commit that added the Apache licence pointed the package metadata at
LICENSE, and the Dockerfile copies only pyproject.toml and README.md before its first dependency
install, so the build stopped with License file does not exist: LICENSE.

The infrastructure could not start. MinIO has withdrawn its Docker Hub repositories, so
minio/minio and minio/mc no longer resolve, and every fresh install stopped at "Starting
infrastructure" with pull access denied. This was not specific to the package: every git clone on a
machine without cached images failed the same way, and it is why the CI test job had been failing. The
machine this project is developed on kept working off copies cached almost two years earlier.

What changed

• The Dockerfile copies the licence before installing, and a test fails if the package metadata ever
  names a file that step does not copy.
• MinIO images come from quay.io. It is the same release and the same image: identical image IDs, and
  the same MinIO commits. The backup and restore scripts, which pulled minio/mc:latest from the
  withdrawn repository, are pinned to the same release.
• scripts/check_images.py resolves every image a fresh install pulls, and the release workflow runs it.
• The release workflow now installs the package on its runner and brings the whole product up before it
  publishes anything, checking the API, the web app, and a request with the generated admin token. The
  v0.1.0 checks only installed the package and ran its launcher, which both defects passed.

Verified for this release

From an empty image cache on a clean Ubuntu 24.04, with no GPU available:

package dependencies	resolved by apt from Ubuntu's own docker.io and docker-compose-v2
first install	about 6.5 minutes
images	api 2.7 GB, web 1.05 GB, database 794 MB
API health	Postgres, Redis and MinIO all ok
web app	answers
authentication	401 without a token, 200 with the generated admin token

sudo apt install ./labeloxav_0.1.1_all.deb
sudo labeloxav install
labeloxav url

What LabeloxAV is

An annotation and data engine for autonomous driving, built around Indian roads: auto-labelling behind a
confidence gate, human review, tracking, 3D from LiDAR or estimated depth, HD map export, and dataset
export in COCO, YOLO, KITTI and OpenLABEL. The package ships the source tree under /opt/labeloxav and a
labeloxav command; the product runs as Docker containers built from that tree.

Honest state: 600,133 labels in the reference corpus, 788 with a human verdict. The HD map runs end to end
but two of its outputs are not yet trustworthy, as the documentation says. Apache License 2.0.

v0.1.0 (superseded: package cannot complete its install)

Do not install this package. It installs and its launcher runs, but labeloxav install fails while building the container images with License file does not exist: LICENSE: the commit that added the licence pointed the package metadata at that file, and the Dockerfile did not copy it before its first dependency install. Found by running the published package end to end on a clean Ubuntu 24.04. A fixed release, v0.1.1, follows with the same contents and a test that prevents this recurring.

The notes below are the original notes for this release and still describe the product accurately.

---

The first tagged release, and the first Debian package.

What LabeloxAV is

An annotation and data engine for autonomous driving, built around Indian roads. It ingests dashcam
video and raw sensor drives, auto-labels behind a confidence gate, routes everything to human review,
tracks objects across frames and cameras, lifts boxes into 3D from LiDAR or estimated depth, measures
accuracy per class with confidence intervals, and exports training data in COCO, YOLO, KITTI and
OpenLABEL. 178 classes, with entries global datasets never had: autorickshaws, hand carts, cattle on the
carriageway.

What the package is, and is not

labeloxav_0.1.0_all.deb installs the source tree under /opt/labeloxav and a labeloxav command.
The product runs as Docker containers built from that tree, so Docker is a dependency and the package
does not turn LabeloxAV into native services.

sudo apt install ./labeloxav_0.1.0_all.deb
sudo labeloxav install        # builds images, starts everything, migrates, creates the admin
labeloxav url                 # http://localhost:3000

No GPU is required for annotation, review, search over existing data, governance and export. The
model paths that need CUDA refuse with a message rather than pretending. A clean install on a machine
with no GPU at all has not yet been tested; that sentence will be removed when it has.

The honest state

• 600,133 labels in the reference corpus, 788 with a human verdict. Most quality numbers are wide.
• The HD map runs end to end for the first time in this release, and two of its outputs are not yet
  trustworthy: fusion can emit a boundary far longer than the drive it came from, and paired lane
  widths run narrow. The documentation says so.
• Four backing services is not a light footprint. The clean-install footprint has not been measured.

Also in this release

• Apache License 2.0. Earlier snapshots were source-visible with no licence.
• A twenty minute narrated tour of every page, with every spoken number read from the live database
  at record time, linked from the README.
• The georeferencer now reads each camera's real calibration. It previously used a nominal lens that
  placed objects 2.5 times too far away near the car and 15 times too far at the horizon.
• Both HD map exports carry lane topology. Previously neither was routable.