WEBVTT

00:00:01.710 --> 00:00:02.845
Hi, I'm Quincy!

00:00:02.845 --> 00:00:05.114
RF-DETR is a real-time object detector.

00:00:05.114 --> 00:00:09.273
It predicts boxes in one shot, and it sizes itself from

00:00:09.273 --> 00:00:10.786
a single training run.

00:00:10.786 --> 00:00:11.542
Here's how.

00:00:13.225 --> 00:00:17.013
The idea: train one network once, then pick any speed-accuracy point

00:00:17.013 --> 00:00:18.391
from it without retraining.

00:00:18.391 --> 00:00:22.180
The named sizes are just points someone chose on one curve.

00:00:23.967 --> 00:00:29.678
An image enters a DINOv2 backbone, a vision transformer pretrained on web-scale images.

00:00:30.098 --> 00:00:33.998
It alternates cheap local windows with a few global layers, then a

00:00:33.998 --> 00:00:35.948
projector hands features to the decoder.

00:00:36.368 --> 00:00:40.231
A fixed set of object queries read the features through deformable attention.

00:00:40.651 --> 00:00:44.723
Class and box heads give each query one label and one box.

00:00:44.723 --> 00:00:45.741
No suppression step.

00:00:45.741 --> 00:00:47.098
It's direct set prediction.

00:00:48.817 --> 00:00:52.232
Normally you train a separate model for every size, and repeat the

00:00:52.232 --> 00:00:53.940
whole search for each new device.

00:00:54.360 --> 00:00:57.420
Here one shared network holds every config at once.

00:00:57.420 --> 00:00:59.460
Each step trains a random one.

00:00:59.460 --> 00:01:02.181
Afterward you grid-search the best, with no retraining.

00:01:03.967 --> 00:01:08.174
On COCO, every RF-DETR size sits above and to the left of

00:01:08.174 --> 00:01:09.576
the other detectors' curves.

00:01:09.996 --> 00:01:12.154
Only the top curve crosses sixty.

00:01:12.154 --> 00:01:16.470
That corner is the largest size, the first real-time detector past sixty.

00:01:18.175 --> 00:01:22.117
Re-measured in one harness, the largest size reaches sixty point one.

00:01:22.117 --> 00:01:27.134
The strong baselines stay below sixty, and even the nano hits forty-eight point four.

00:01:28.767 --> 00:01:31.915
Train one shared network, then pick your operating point.

00:01:31.915 --> 00:01:36.113
That is how a detector family crossed sixty AP in real time.

00:01:37.692 --> 00:01:42.645
To recap: a DINOv2 backbone, one-shot set prediction, one shared network for

00:01:42.645 --> 00:01:46.361
every size, and the first real-time detector past sixty.

00:01:46.781 --> 00:01:48.908
Every source is in the full article.

00:01:48.908 --> 00:01:49.516
I'm Quincy.

00:01:49.516 --> 00:01:49.820
Bye!
