@gamechangerai: A road infrastructure client was bidding for a city project against vendors promising cameras and image recognition. They had sensors in the ground and no vision model at all. I got two sprints of two weeks, final presentation included. The first decision was to train on the categories the buyer already files. Longitudinal cracks, transverse cracks, alligator cracks, faded lane lines, faded crosswalks and manhole covers. ⠀ I did not invent a taxonomy, because matching the customer vocabulary is what makes the output usable on the day it lands. The second decision was that detection alone is not a deliverable. Every image carried GPS metadata, so I plotted each detection on a map, colour coded by damage type, with a checkbox filter per category and a timestamp on each point. The people in that room read a map and a filter. A confusion matrix would not have moved the conversation one inch. ⠀ The third decision was to keep the stack boring on purpose. A standard object detection model fine tuned on images the client already had, a common inference server pulling weights from object storage, and a lightweight web front end behind one API. Nothing on that list was novel, which is exactly why it was finished in time to present. I also handed over the retraining, so they could add a new damage category without calling me. What would you cut first if your demo was in four weeks? ⠀ Follow @gamechangerai for more. ⠀ #ai #machinelearning #computervision #objectdetection #smartcity

606

There are no more comments for this video.

To see more videos from user @gamechangerai, please go to the Tikwm homepage.