3 ms·
And yet, most of the time, humans don’t seem to require radar to distinguish the moon (or a big yellow light far away) from a traffic light. I think stereo vis
by Someone 5y ago
And yet, most of the time, humans don’t seem to require radar to distinguish the moon (or a big yellow light far away) from a traffic light.
I think stereo vision won’t help humans much at the distances we’re talking about (might be different for a car with cameras over a meter apart, but that will suffer from bad resolution compared to the human fovea).
Parallax probably is a clue we use. You want to brake gently to optimize comfort for the passengers, so let’s say you want to start braking 10 seconds away from the traffic light. Since sin(x) ≈ x, the image of the traffic light will move up about 10% every second at that moment.
On a camera, that could be less than 100 pixels, but I think that should be detectable.
⇒ you don’t have to discriminate between the moon and a tragic light, you just have to find the nearby yellowish blobs.
Is it publicly known what Tesla’s self-driving does? Given the above, I think I would try to extract a depth field from the known speed of the car and the camera images, discard or deprioritize whatever is more than 20 seconds out, and then let some ML algorithm loose on detecting points of interest such as traffic lights, cars, etc. (As opposed to “feed a ML algorithm lots of video, and hope it will learn to deprioritize stuff that’s far away)
- gerash 5y agoNeural nets esp. those than can run several times a second on a car (not in datacenters) are still very far from human level of perception. A human disambiguates by looking around, lots of prior knowledge about how the world works, etc. That doesn't seem to be the case with the current Teslas.