WEBVTT

00:00:01.710 --> 00:00:02.729
Hi, I'm Peregrine!

00:00:02.729 --> 00:00:06.466
Let's see how Oren maps the distance to every surface, continuously,

00:00:06.466 --> 00:00:07.824
and in real time.

00:00:09.892 --> 00:00:14.237
The octree carries the structure, so the network only learns a small correction.

00:00:14.237 --> 00:00:17.913
That keeps it fast, and keeps the distance field accurate everywhere.

00:00:20.425 --> 00:00:22.057
A sensor streams in points.

00:00:22.057 --> 00:00:25.974
Oren drops them into a semi-sparse octree and stores a distance and

00:00:25.974 --> 00:00:27.606
a gradient at every corner.

00:00:28.026 --> 00:00:32.390
Interpolating those corners, with the gradients, gives an explicit distance prior that

00:00:32.390 --> 00:00:34.935
is accurate even far from any surface.

00:00:35.355 --> 00:00:38.505
The same corners also hold a few tiny learned features.

00:00:38.925 --> 00:00:43.212
A small network, just two narrow layers, reads a feature together with the prior.

00:00:43.632 --> 00:00:45.623
It adds a small residual.

00:00:45.623 --> 00:00:50.002
Prior plus residual is the final distance, continuous and never truncated.

00:00:50.422 --> 00:00:54.035
Because the octree holds the structure, the network stays small and does

00:00:54.035 --> 00:00:55.240
not forget the scene.

00:00:56.692 --> 00:01:00.845
The usual truncated map only knows distances right next to a surface.

00:01:01.265 --> 00:01:04.555
Oren keeps the true distance everywhere, so a planner can read its

00:01:04.555 --> 00:01:05.651
clearance in the open.

00:01:06.925 --> 00:01:11.834
Across the whole room, Oren's mean distance error is two point one five centimetres.

00:01:11.834 --> 00:01:13.236
Both baselines are worse.

00:01:14.300 --> 00:01:18.061
On the large outdoor sequence it runs at over fourteen frames a

00:01:18.061 --> 00:01:21.508
second, the fastest tested, on about one and a half gigabytes.

00:01:22.883 --> 00:01:25.191
Here are the distance fields as a slice.

00:01:25.191 --> 00:01:27.788
Oren fills the whole room, matching the ground truth.

00:01:28.208 --> 00:01:32.028
The truncated methods only colour a thin band near surfaces, and leave

00:01:32.028 --> 00:01:33.302
the open space blank.

00:01:34.975 --> 00:01:38.600
A good explicit prior leaves the network almost nothing to learn.

00:01:40.942 --> 00:01:44.990
Octree prior, tiny residual, a full distance field in real time.

00:01:44.990 --> 00:01:45.726
That's Oren.

00:01:46.147 --> 00:01:48.283
Every source is in the full article.

00:01:48.283 --> 00:01:48.893
I'm Peregrine.

00:01:48.893 --> 00:01:49.198
Bye!
