Production-style computer vision pipelines, built one demo at a time and in public.
Each demo takes a real problem from industry, decomposes it into components with explicit data contracts, grounds every model choice in the research record, and measures every phase against a bar written down before the run. The goal is to learn by building things the way a production CV team would, not by stitching tutorials together.
| # | Demo | Problem | Status |
|---|---|---|---|
| 01 | Parked vs Stopped vs Moving | Classify every vehicle seen from a moving bus camera as parked, stopped or moving, for bus-lane and double-parking enforcement. Parked and stopped both have zero speed; only context and duration separate them, so this is temporal intent, not motion. | Phase 1 data extracted (850 scenes, 34k tracks); label spot check and Phase 2 next |
- Decompose before building. Each component answers one question and hands a defined table downstream. The decomposition is written down before any model is picked.
- Ground every model choice in the research record. Papers from the source, the standard datasets, the leaderboard and its metric. References live in the demo's PLAN.md.
- Baseline first, then learn. Every learned model must beat a simple, documented baseline on the same data and metric.
- Pre-register and measure. The pass/fail bar is written before the experiment runs and never moved after. Metrics are reported per component and per failure class.
- Verify visually. Every metric ships with montages, clips or plots of the cases behind it.
Each demo folder has the same shape, described in docs/DEMO_TEMPLATE.md:
a PLAN.md that is the source of truth, one package per component with typed contracts
between them, an experiments/ folder with a pre-registered bar per run, and a --selftest
on every runnable module that works with no GPU, no dataset and no network.
.
├── CLAUDE.md working rules and conventions for anyone (or any agent) working here
├── docs/ templates: demo layout, experiment pre-registration
├── packages/cvcommon/ shared: table contracts, selftest harness, logging
└── demos/
└── 01_parked_stopped_moving/
Requires uv and Python 3.11 or newer.
uv sync # creates .venv with every package in the workspace
make check # ruff + pytest + every module's --selftestDatasets are never committed. Each demo's README says what to download and where to put it;
nothing in make check needs any of it.
MIT. See LICENSE.