frame-lab
ActiveA LeetCode-style practice lab for perception algorithms: small, testable Python exercises for the geometry and detection math everyone uses and few implement.
Implementation fluency, on purpose
Most perception knowledge lives inside large libraries: you call nms(),
you trust it, and the actual algorithm quietly atrophies in your head.
frame-lab is my counter-practice: a local, LeetCode-style repo where each
perception primitive is a self-contained problem folder with a README, a
starter file, a reference solution, and a test suite.
The curriculum covers the math that shows up in every perception interview and every real pipeline: bounding-box geometry and IoU, non-maximum suppression, coordinate-frame transforms and projections, tracking and association utilities, sensor-data handling.
How a problem works
- Read the problem statement.
- Implement it in
starter.py. - Run the tests:
uv run pytest problems/001_.../test_starter.py. - Stuck? The reference
solution.pyis right there; this is practice, not an exam.
Reference solutions are validated in CI, so the answer key itself is tested.
Why it earns a spot here
It’s a practice repo, but it’s also the sharpening stone for everything else on this page: the transform conventions in slam-core, the geometry in sfm-core, the track handling in msckf-vio all got faster to write because the primitives were rehearsed here first. New problems land whenever I catch myself reaching for a library call I couldn’t re-implement cold.