Vision scratch experiments

Problem

Before the vehicles, there was the question: how does a computer see anything at all? The problem was not a product or a competition — it was apprenticeship. I needed to learn to use a computer to see stuff, and the way to learn was to build the smallest possible experiments.

Difficulty

The difficulty was that nothing worked the first time, and there was no vehicle to blame. Face regeneration, color-sign detection, tracking, prediction — each experiment failed in its own way, and each failure was the actual curriculum.

Built

A set of scratch experiments, each a small repo: Face-Regeneration (learning to use a computer to see stuff), Detect-color-signs, Track_and_Predict, and decision-tree-classifier. All four were pushed the same day (2023-11-29) — a single-day learning burst. Small, honest, and built from scratch.

Owned

All of it — the experiments were mine, the failures were mine, and the apprenticeship was mine.

Failed

The experiments failed constantly, which was the point. The failures were not setbacks; they were the mechanism. Each one showed a piece of how vision actually works.

Changed

The phase changed what I could build. The experiments became the foundation for the perception work that came later — the Duburi interview task and the subteam work grew directly out of this apprenticeship.

Machine-now

Retired. The phase is over and the experiments are archived, but the apprenticeship they paid for is still running.

Lesson

Decision tree classifiers are awesome and scary — and so is everything else in vision when you build it from scratch. The origin phase taught me that the way to understand a machine is to build it badly first.