Thread by @rice_fry

Tesla has added new voxel 3D birdseye view outputs and it's pretty amazing!

Nice of them to start merging some bits of FSD into the normal firmware in 2021.36 so we can play with the perception side 🙂

Thanks to @greentheonly for the help!

Most of the critical FSD bits are missing in the normal firmware. These outputs aren't normally running but with some tricks we can enable it.

This seems to be the general solution to handling unpredictable scenarios such as the Seattle monorail pillars or overhanging shrubbery.

The nets predict the location of static objects in the space around them via a dense grid of probabilities.

The output is a 384x255x12 dense grid of probabilities. Each cube seems to be ~0.33 meters and currently outputs predictions ~100 meters in front of the vehicle.

This is similar to the previous single camera depth model but given it the birdseye view treatment.

See our previous tweets on that at:

https://twitter.com/greentheonly/status/1412597331984367616

Here's an example of a post in the middle of a narrow road. fn.lc/s/depthrender/…

Before this Tesla would have to manually label this as part of the training set to ensure the car doesn't run into it

Here's a full intersection, outputs seem quite reasonable in all directions. You can see the 4 buildings on each side, the curbs ahead as well as the trees by the side of the road.

fn.lc/s/depthrender/…

Here's a hard median with a post that it correctly identifies. Adds another level of safety to ensure that the car doesn't drive over hard curbs.

This example highlights that the model ignores cares and only shows the static objects. fn.lc/s/depthrender/…

Pretty impressive how much detail it can capture of trees and the landscaping on the side of the road.

fn.lc/s/depthrender/…

Here's the view leaving a parking lot. Clearly distinguishes where the road is vs the T-bone style intersection.

fn.lc/s/depthrender/…

You can check out the raw data at fn.lc/s/depthrender/…

And the corresponding time synced video is at

The uploaded voxel frames are every half second for practicality reasons (in the car it's much higher FPS)

I suspect they're taking the same offline 3D models they use to label the birdseye view training data (as seen during AI day) and converting it to voxel data to train a net.

It's a very clever solution, kudos to the engineers who worked on this.

I'm very curious what the model architecture looks like and how much it differs from the other birdseye view nets.

The 3D convolutional NNs used here are similar to what could potentially be used merge radar with vision if Tesla can get access to the raw Conti radar data.

The 3D birdseye view is a fair bit lower resolution than LIDAR but very impressive and achieves much of the same purpose.

These models are outputting probabilities so you can see where the model is confident vs not.

I don't quite know what the scale is here but having a 75% threshold seems to work pretty well. For all these renders I only show voxels that are above the target threshold

Rendering voxel data in a browser is pretty tricky so if anyone wants to help with more advanced visualizations let me know :)

I see why Elon said they were having issues visualizing it