Fold Step Finder

What this does. Upload a video of a person folding a T-shirt. Stage 1 checks the footage is usable (plays, not frozen or black, sharp, well lit). Stage 2 marks when each step happens: grab, spread out, fold, put in corner.

How stage 2 works. Every half second a frozen image model (CLIP) describes the frame and the frames 1 s either side; a small classifier trained on 10 human-labelled clips scores the four steps; a decoder smooths the result into a timeline (steps mostly go forward, a single odd moment doesn't start a new step).

How good is it. On 20 EgoDex clips it never saw, it named the same step as a human reviewer 80% of the time, with step starts typically 0.8 s apart. It has only been trained on EgoDex footage; other camera setups need a few labelled clips first.

Allow 15 to 60 seconds per clip, depending on length and on whether the app was asleep.

Stage 1 · Footage check

Stage 2 · Steps found