Continued efforts to build out better detection and annotation tools. Using SAM3, creating first pass annotations, then validating before kicking off additional active learning.

Created a "swipe" feature set to swipe to approve/reject/skip annotations.

Playing with different annotation interactions, attempting to improve speed and comfort.

Now using a couple ML methods to detect contours, then be able to trim or add to the polygon with limited interaction.

Probably need to make an iOS app with pencil support...

Experimenting with animating sensor data. Eventually looking to build out predicted poses based on sensor data, and how that maps to detected activities.

Boring, for now.

I've been thinking of new ways automated testing could be used. Today, I created a series of tests that goes through my customer flows in a variety of appearances.

They take screenshots, finally creating a new webpage where I can directly compare various UI against each appearance.

Been doing a ton of annotation work in Clay lately and noticed something fun.. corner-to-corner bounding boxes just feel better than clicking edges.

At first I thought it was just fewer clicks, but it's more than that. You're really just answering two simple questions: where does this thing start? Where does it end?

Edge-clicking makes you think about width, then height..your brain has to decompose the object instead of just seeing it. Hundreds of annotations in and my brain is way less fried