4 ms·
Right? This should be trivially solved with some sort of temporal filtering on detected objects. If you detect a traffic light at (x,y) in one frame, and it dis
by ModernMech 5y ago
Right? This should be trivially solved with some sort of temporal filtering on detected objects. If you detect a traffic light at (x,y) in one frame, and it disappears in the next, but there's a new one at (x+dx,y+dy), then you shouldn't place a new one down in the world frame. You should only place a traffic light down if you're confident it exists and is operational. At the very least, the lights should be detected on the back of the truck, but they should move with the truck. At least that matches what's happening.
I don't understand why this is hard for Tesla engineers -- I was doing this kind of thing in grad school a decade+ ago and it worked fine. I've seen it in other demos where object classifications rapidly cycle between person, bike, car, etc. Are they not filtering anything? Is this a symptom of "AI-ing all the things"? Because we did it with bog standard computer vision techniques back then and never got behaviors like this.
- nmca 5y agoDid any of your grad school work make it into the real world at all? Typically I'd suggest that such ideas are simple in theory and difficult in practice.
- ModernMech 5y agoYes, there are real systems out there working off of the techniques we used back then. I didn't work much in theory at all.