WEBVTT

00:00:01.290 --> 00:00:02.179
Hi, I'm Calloway!

00:00:02.179 --> 00:00:06.329
Two systems rebuild the world from a moving camera, one frame at a time.

00:00:08.392 --> 00:00:11.844
Both freeze the same depth model, and stop storing a dense

00:00:11.844 --> 00:00:13.099
point map per frame.

00:00:13.099 --> 00:00:15.296
They keep a small bounded summary instead.

00:00:16.842 --> 00:00:21.822
Each video frame enters a frozen geometry model that already knows depth and shape.

00:00:22.242 --> 00:00:26.192
It turns the frame into one small token, a compact summary of the view.

00:00:26.612 --> 00:00:30.584
A tiny head compares tokens in pairs and predicts camera motion, scoring

00:00:30.584 --> 00:00:32.571
how much to trust each one.

00:00:32.990 --> 00:00:37.146
The trusted links fold into one running estimate, keeping only a fixed

00:00:37.146 --> 00:00:38.531
set of key frames.

00:00:38.531 --> 00:00:39.570
Memory stops growing.

00:00:41.025 --> 00:00:43.556
The old way pins everything to frame one.

00:00:43.556 --> 00:00:47.353
Travel far, and the numbers grow without bound, so the track drifts.

00:00:47.773 --> 00:00:51.931
The first system predicts only motion between pairs, weighted by confidence, so

00:00:51.931 --> 00:00:54.010
any frame can anchor the next.

00:00:55.633 --> 00:00:59.608
The second system tracks each frame fast against a tiny local memory.

00:01:00.028 --> 00:01:03.271
Chunks of the map are stitched by near links, longer links, and

00:01:03.271 --> 00:01:04.893
loop links when a place returns.

00:01:05.313 --> 00:01:08.947
It all becomes one pose graph, with no slow bundle solve, so

00:01:08.947 --> 00:01:10.764
moving cars no longer break it.

00:01:12.467 --> 00:01:16.486
On streaming benchmarks the first system has the lowest trajectory error, and

00:01:16.486 --> 00:01:18.161
it is the smallest model.

00:01:19.717 --> 00:01:23.096
Add LiDAR on a brutal highway drive, and the error falls from

00:01:23.096 --> 00:01:24.786
fifty six metres to under three.

00:01:25.975 --> 00:01:28.391
The geometry model is now a frozen part.

00:01:28.391 --> 00:01:32.014
The real work is holding memory flat and skipping the slow solve.

00:01:34.817 --> 00:01:39.016
Freeze the depth model, keep a bounded summary, trust each link, and

00:01:39.016 --> 00:01:40.766
stream geometry from endless video.

00:01:41.186 --> 00:01:43.252
Every source is in the full article.

00:01:43.252 --> 00:01:43.842
I'm Calloway.

00:01:43.842 --> 00:01:44.137
Bye!
