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

mikes_fsd

Banned
May 23, 2014
2,562
2,684
Charlotte, NC
I propose we change the definition of a HD map to: "A Map that has been created prior to an autonomous vehicle can drive the area"
This is a bad definition because this makes conventional nav maps hd maps (obviously you must have nav maps to navigate in an area)

Just like in 2005 today's maps would be considered HD.

I am going to offer up the following for a definition that separate HD Maps (hard to maintain) vs High Precision Map (as proposed by Tesla in 2015, today is just normal map data in your Google Maps).
If maps require any pre-rendered tiles of surroundings to be fed into the NN, they are HD maps. They are used as foundation for the NN's processing. They take time to capture, process and maintain on top and are a separate feed from the navigational maps.

Hight Precision Maps are not pre-rendered maps, they are fed into the NN's and are used in conjunction with other feeds but are not the baseline for the output. The same map data is used for navigation as well.

Hope it could work as a delineator in the definitions.
 
Last edited:

ZeApelido

Active Member
Jun 1, 2016
3,300
29,803
The Peninsula, CA
Maps is the wrong term to use I guess.

Waymo creates a full 3D reconstruction of every road, offline, that then is available to it's cars to compare against real-time 3D reconstruction.

Tesla probably feeds in some basic map knowledge (coordinates of intersection objects, street structures?) of roads and intersections into its real-time algorithm for 3D reconstruction.

Tesla does not create a full 3-D reconstruction of the environment offline for comparison.
 
  • Like
Reactions: mikes_fsd

MP3Mike

Well-Known Member
Feb 1, 2016
16,422
38,301
Oregon
As it should be and as Tesla has stated that they use maps.
How is navigation supposed to be done in a FSD car if the car does not have maps?

The argument is not on whether maps are used, the argument is that it is not HD Maps (not highly detailed cm level accuracy maps)

In 2015 a full year before Tesla revealed the Autopilot 2.0 hardware.. there was this article Tesla Building Next Gen Maps through its Autopilot Drivers
How Tesla was getting fleet data to improve the accuracy of maps. At that time not many providers had lane level info for navigation (Google maps had just introduced the lane level info in their turn-by-turn navigations)
Tesla-Map-Comparison.jpg

This does not give you lane marking drawn on tiles.
It gets you info along the lines of: 4 lanes in one direction off-ramp starts at this geo-code, etc

Just remember that this mapping effort was one of the things that delayed FSD, Elon said that that got them to "local maximums" and that they had to abandon the high precision mapping and go back to relying on vision to make further progress.

Also, by definition anything that requires pre-existing maps, of any resolution, can not ever be L5. Because L5 has to be able to drive on a brand new freshly paved, or dirt/gravel, road. And we have seen examples of the Tesla beta FSD driving on roads that aren't even in Google Maps yet.
 

mikes_fsd

Banned
May 23, 2014
2,562
2,684
Charlotte, NC
Just remember that this mapping effort was one of the things that delayed FSD, Elon said that that got them to "local maximums" and that they had to abandon the high precision mapping and go back to relying on vision to make further progress.

Also, by definition anything that requires pre-existing maps, of any resolution, can not ever be L5. Because L5 has to be able to drive on a brand new freshly paved, or dirt/gravel, road. And we have seen examples of the Tesla beta FSD driving on roads that aren't even in Google Maps yet.
Oh I know that.
They realized Maps (HD or otherwise) cannot be the foundation. Vision has to be the foundation and map data is just another input that enhances the final output.
 
Last edited:

Mardak

Member
Oct 13, 2018
807
1,648
USA
so what accuracy is there on the tesla maps for traffic control devices, crosswalks, stop lines and other such objects?
At least looking at OpenStreetMap data, intersections with traffic lights and stop signs often have a single node (e.g., the point where two road lines cross) with an attribution tag indicating there is a given road feature. If we just assume most intersections are of 2 lanes in each direction with roughly 5 meters width for each lane, the sign or signal most likely is along the edge or corner of the intersection (and not in the exact middle of the road), so the accuracy is probably around 5-10 meters for a small intersection. And this is based on some human overlaying a line on top of an ideally somewhat reasonably GPS oriented satellite image.

Another way to think about this is the actual data structures needed to encode OSM/Tesla map data vs "HD maps." From a quick search, HD maps generally have relatively precise 3D positioning and adjacency data for every single lane subsegment and sign/signal whereas OSM 2D vector map data has points, lines and areas that have coarse locations with tags. It's not uncommon in OSM for a 4-lane road to be identified with a single line with an attribution of lanes=4 vs a HD map would have information about each of the 4 lanes including where the lane gets wider or narrower.
 
  • Like
Reactions: NHK X and mikes_fsd

mt09

Member
Dec 11, 2018
30
71
IL

I'm interested in knowing what's being compared in this exchanging of outputs. There's so much noise watching the visualization with edges jumping around that you'd think a disagreement would occur quite frequently. Both road features and path planning (sometimes flipping from right/left turn) seem unstable.

It's possible this is not enabled yet, but I was also hoping the rewrite fourth dimension (time) would add some stability to features disappearing.
 
  • Like
Reactions: J1mbo

mikes_fsd

Banned
May 23, 2014
2,562
2,684
Charlotte, NC
can maps be not-prerendered?
Vector tiles - Wikipedia
Vector tiles, tiled vectors or vectiles are packets of geographic data, packaged into pre-defined roughly-square shaped "tiles" for transfer over the web. This is an emerging method for delivering styled web maps, combining certain benefits of pre-rendered raster map tiles with vector map data. As with the widely used raster tiled web maps, map data is requested by a client as a set of "tiles" corresponding to square areas of land of a pre-defined size and location. Unlike raster tiled web maps, however, the server returns vector map data, which has been clipped to the boundaries of each tile, instead of a pre-rendered map image.
 

ZeApelido

Active Member
Jun 1, 2016
3,300
29,803
The Peninsula, CA
how can maps be not-prerendered? Maps is a prior knowledge you have of the area before you actually enter the area.

You might want to rethink your definition a bit.

They are not creating a 3D environment offline.

My guess is they are taking some simplistic information about the intersection, and predicting what it should look like based on learnings of intersections in general. Thus the output looks quite realistic, though it is actually mostly a prediction.
 

verygreen

Curious member
Jan 16, 2017
2,945
11,510
TN
yes, that's what Tesla does. They use valhalla/mapbox container with tiles.

Millimeter precision HD Vector Maps (the title of that page is fun btw)

Here's Tesla fork of the code in their own github: teslamotors/valhalla

This commit is support for loading gzipped tiles: support loading gzipped tiles · teslamotors/[email protected]

they then feed these vectors and other features along the route into the NN.

So by not pre-rendered you mean vector tiles? You should have used this term then.

flat fixed resolution "pixel" tiles are useless anyway - you cannot easily scale them you cannot know easily what's on them without machine vision (that needs to basically convert it back into the vector form) and so on. Ican't think of anybody doing this.

Now I am not sure you can call lidar 3D world scans "pre-rendered"/"raster" because I have no idea how they actually store that stuff.

They are not creating a 3D environment offline.
They don't. We did not really discuss anything about 3D in prior discussion, why bring in 3D in all of a sudden?

My guess is they are taking some simplistic information about the intersection, and predicting what it should look like based on learnings of intersections in general. Thus the output looks quite realistic, though it is actually mostly a prediction
Yes, this is what they do. Most importantly, if something is not in that "simplistic" information set as you call it (like say there's no side street marked in there) - the system does not see it until it's right next to it and even then it's not very realistic (see the SC road link I posted for examples) to have actual high confidence visual. When something IS marked in the "simplistic information" set - the system "predicts" it based on a lot lower confidence visuals.

I’ve been following along, and at this point, you’re coming off as purely a contrarian with comments like this.
huh? what's contrarian about it? I am just trying to get a non-contradictory definition.
 

Products we're discussing on TMC...

About Us

Formed in 2006, Tesla Motors Club (TMC) was the first independent online Tesla community. Today it remains the largest and most dynamic community of Tesla enthusiasts. Learn more.

Do you value your experience at TMC? Consider becoming a Supporting Member of Tesla Motors Club. As a thank you for your contribution, you'll get nearly no ads in the Community and Groups sections. Additional perks are available depending on the level of contribution. Please visit the Account Upgrades page for more details.


SUPPORT TMC