visiond: replace wiggly model with non wiggly model

[geohot]

Retrained the temporal model on the same dataset but added more dropout. This fixes the wiggle issue, potentially at the expense of responsiveness. Comment here on which model you like the best and issues with them:

1. 0.5.3 model
2. 0.5.4 model
3. 0.5.4-nowiggle model

[ch4se]

Wiggles are better because you can see that OP is actually working hard at making decisions

[jfrux]

I tend to agree Chase. I do think we could maybe do something with it… but I’m not sure how we could get the best of both worlds.

…

On Sep 27, 2018, 1:05 PM -0400, Chase Higgins ***@***.***>, wrote: Wiggles are better because you can see that OP is actually working hard at making decisions — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.

[geohot]

From a purely visual perspective, I would not say wiggles inspire confidence.

[openmechanics]

1. • 0.5.3 model-Was cool
2. • 0.5.4 model-Drives smooth better then previous models however the only down side is it's trying to be a detective lol best way to describe it. It's trying to find missing lanes and once it finds them it's trying to go that way. So on streets when the lane ends it wants to go full left except for straight.
3. • 0.5.4-nowiggle moel- Haven't tried yet.

[ch4se]

Yeah I wouldn’t say wiggles inspire confidence, but it at least gives you an idea of when the model actually is confident.

Wiggly model probably makes the user pay more attention, people seem to be overly confident in the system.

[jfrux]

I’m sure I don’t understand it well enough so forgive me, but could the line not be rendered two different ways? Is the line rendering itself REALLY tied to the model itself and not to the renderer (not sure if thats a thing… or if its not, it SHOULD BE :p) Maybe some sort of `ModelRenderer` class that can decide some of those properties? Or is it literally burned into the logic and cannot be swayed with a switch or toggle?

…

On Sep 27, 2018, 1:29 PM -0400, Chase Higgins ***@***.***>, wrote: Yeah I wouldn’t say wiggles inspire confidence, but it at least gives you an idea of when the model actually is confident. Wiggly model probably makes the user pay more attention, people seem to be overly confident in the system. — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.

[openmechanics]

Yeah I wouldn’t say wiggles inspire confidence, but it at least gives you an idea of when the model actually is confident.

Wiggly model probably makes the user pay more attention, people seem to be overly confident in the system.

agree^ requires user to pay more attention:)

[jeankalud]

Tested both 0.5.4 with and without wiggle. I prefer the no_wiggle version personally.

I liked seeing the wiggle version trying to find the lines but it doesn't inspires confidence. Also, compared to the previous 0.5.3 version I feel it should be given less authority when confidence level is low. Right now it tends to turn right-left back and forth in some no lines conditions. This should be smoothed out imho.

I REALLY appreciate the increased turn angle limit, its amazing, so manu thanks for this.

[zeeexsixare]

Both versions of 0.5.4 drive similarly for me. I agree that wiggles don't inspire confidence. However, the white lines , or the green path line could potentially be turned orange or red dynamically if we want to "feel" how confident the model is. This way we can remove the wiggles but retain model confidence.

I also agree with jeankalud that visiond being able to steer to greater extremes is really nice when the model gets it right. It successfully navigated a low-speed curve that I had to manually override before.

[pjlao307]

no-wiggle feels more "natural" to me on surface streets. On highway it seems very similar.

[ewumnrr]

Works well, lane detection is more stable. All in all an improvement on local roads and gives very solid highway performance.

[jfrux]

RIP