Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

What's With The Dancing Cars???

This site may earn commission on affiliate links.
I think it's most likely sonar reflection which isn't as likely when your car and other cars are moving around. They must be combining the sonar and camera feeds to get the size, position and 'facing' of the vehicle. With reflections, I could easily see how the cars could dance around. I also suspect as the algorithm for the sensor fusion improves again that it will stop happening as much.
 
I think it's just a UI glitch, and not actually how AP is interpreting the vehicles. You can see "AP-vision" on a lot of the tweets by @verygreen green (@greentheonly) | Twitter (my apologies if I tagged the wrong user).

My personal theory: the UI only ever tries to display vehicles parallel to lane lines (as evidenced by cross-traffic moving sideways when you're stopped at a red light). The UI is either getting confused by obscured lane lines behind the car, or it's applying some sort of "angle-tweaking" function in a way that spirals out of control.
 
It's pretty simple actually. When moving, the radar system can calculate the velocity (speed + direction) of all the cars near you pretty easily. So when it renders the car, it renders which way it faces.

When standing still or at low speeds, the radar velocity isn't really a thing. So the cameras also calculate velocity of cars. You'll find that the dancing is very close to the frame rate of the camera.

Since the speed of the cars is very low, it's really easy to get velocity errors. This includes what direction the car is travelling, so it looks like it dances a bit.

Expect this to improve with HW3 + more NN training.
 
It's pretty simple actually. When moving, the radar system can calculate the velocity (speed + direction) of all the cars near you pretty easily. So when it renders the car, it renders which way it faces.

When standing still or at low speeds, the radar velocity isn't really a thing. So the cameras also calculate velocity of cars. You'll find that the dancing is very close to the frame rate of the camera.
This is also why the cars behind you where the most powerful radars are located are dancing the most. not.
 
This is also why the cars behind you where the most powerful radars are located are dancing the most. not.

Please learn to read? I specifically said the cars that dont have good information for velocity (aka arent measured by the radar counts too) are the ones that dance.

Behind you is the one repeater camera and that's it. Makes sense that at a stop velocity direction bounces around
 
Please learn to read? I specifically said the cars that dont have good information for velocity (aka arent measured by the radar counts too) are the ones that dance.

Behind you is the one repeater camera and that's it. Makes sense that at a stop velocity direction bounces around
The ruth is they all dance, back and front.

Not sure about your car, ut on my car there are three cameras looking back: two repeaters and the backup cam. Velocity does not play a role there all that much either, it's the car orientation detection by visual means. there are plenty of examples of cars doing 360 degree rotations driving by.
 
The ruth is they all dance, back and front.

Not sure about your car, ut on my car there are three cameras looking back: two repeaters and the backup cam. Velocity does not play a role there all that much either, it's the car orientation detection by visual means. there are plenty of examples of cars doing 360 degree rotations driving by.

Any intuition on how the orientation is being calculated?

I imagine it's not doing the equivalent of "Does this end look like headlights? Does this end look like taillights?" but rather "Cars are longer than they are wide, so the short side of the rectangular prism headed towards me must be the front." So the 360 spins could be explained by the system seeing a shape approaching it, estimating the length and width, and that estimate changing as more information is added.
 
I've had wiggling, dancing, even 360 degree spinning cars while stationary in line at Starbuck's drive thru. In fact that's probably where I see the most of them. Slowing down to a stop in traffic at a light would be second. Also ghost vehicles or ghost people show up sometimes too. Does make you wonder under what conditions they appear.

Could also mention that the majority of the dancing cars tend to be in front of me or to a lessor degree along side of me. I've also wondered if the movement isn't influenced by the music selection I have on. Sure seems so sometimes :).
 
Any intuition on how the orientation is being calculated?

I imagine it's not doing the equivalent of "Does this end look like headlights? Does this end look like taillights?" but rather "Cars are longer than they are wide, so the short side of the rectangular prism headed towards me must be the front." So the 360 spins could be explained by the system seeing a shape approaching it, estimating the length and width, and that estimate changing as more information is added.
NN outputs orientation (in the form of a 3D bounding box with different car sides clearly marked as front, back left and right) based on camera images.