DETR: End-to-End Object Detection with Transformers | Paper Explained

preview_player
Показать описание

In this video I cover DETR, an end-to-end object detection pipeline with transformers.

The main 2 ideas are:
* Using transformers instead of specialized vision architectures
* Using Hungarian matching and loss to train the system e2e

▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬

⌚️ Timetable:

00:00 Intro: DETR main ideas
00:45 Non-max suppression
03:20 High-level pipeline overview
07:50 Architecture in more detail
12:10 Matching loss
18:35 Hungarian loss
21:00 Results
24:05 Visualization
27:35 Ablations
30:00 Outro: Segmentation results

▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
💰 BECOME A PATREON OF THE AI EPIPHANY ❤️

If these videos, GitHub projects, and blogs help you,
consider helping me out by supporting me on Patreon!

One-time donation:

Much love! ❤️

Huge thank you to these AI Epiphany patreons:
Eli Mahler
Petar Veličković
Zvonimir Sabljic

▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬

💡 The AI Epiphany is a channel dedicated to simplifying the field of AI using creative visualizations and in general, a stronger focus on geometrical and visual intuition, rather than the algebraic and numerical "intuition".

▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
👋 CONNECT WITH ME ON SOCIAL

👨‍👩‍👧‍👦 JOIN OUR DISCORD COMMUNITY:

📢 SUBSCRIBE TO MY MONTHLY AI NEWSLETTER:

💻 FOLLOW ME ON GITHUB FOR COOL PROJECTS:

📚 FOLLOW ME ON MEDIUM:
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬

#detr #objectdetection #transformers
  1. Aleksa Gordić - The AI Epiphany
  2. detr
  3. object detection
  4. segmentation
  5. object detection with transformers
  6. transformers
Рекомендации по теме
DETR: End-to-End Object 0:40:57

DETR: End-to-End Object Detection with Transformers (Paper Explained)

DETR - End 0:09:35

DETR - End to end object detection with transformers (ECCV2020)

DETR: End-to-End Object 0:31:19

DETR: End-to-End Object Detection with Transformers | Paper Explained

DETR: End-to-End Object 0:06:14

DETR: End-to-End Object Detection with Transformers

Object Detection Part 0:04:28

Object Detection Part 7: Detection Transformers (DETR), Object Queries

Decoding DETR 1:08:51

Decoding DETR

DETR - End-to-End 0:02:02

DETR - End-to-End Object Detection with Transformers - DEMO

[Tutorial] Training End-to-end 0:25:16

[Tutorial] Training End-to-end Object Detection with Transformer(DETR) model on custom dataset

DETR End to 0:26:55

DETR End to End Object Detection with Transformers

Reimplementation Showcase DETR: 0:01:07

Reimplementation Showcase DETR: End-to-End Object Detection with Transformers | CVDL Final Project

DETR (DEtection TRansformer) 0:15:08

DETR (DEtection TRansformer) | Lecture 38 (Part 2) | Applied Deep Learning (Supplementary)

[Paper Review] DETR: 0:15:36

[Paper Review] DETR: End-to-End Object Detection with Transformer

Facebook DETR | 1:42:52

Facebook DETR | ML Coding Series | End to end object detection with transformers

Object detection Using 0:18:21

Object detection Using Detection Transformer (Detr) on custom dataset

VPS-01: DETR: End-to-End 0:24:04

VPS-01: DETR: End-to-End Object Detection with Transformers paper review

DETR: End to 0:03:21

DETR: End to End Object Detection and Panoptic Segmentation

DETR - End-to-End 0:02:53

DETR - End-to-End Object Detection with Transformers - DEMO

PR-284: End-to-End Object 0:34:56

PR-284: End-to-End Object Detection with Transformers(DETR)

DINO - DETR 0:41:10

DINO - DETR with Improved DeNoising AnchorBoxes for End-to-End Object Detection

Experimenting Object Detection 0:11:24

Experimenting Object Detection with DETR

Object Detection with 0:09:01

Object Detection with Transformers | DETr | PyTorch

ELI5: DETR 0:01:52

ELI5: DETR

Deformable DETR | 0:08:17

Deformable DETR | Lecture 38 (Part 3) | Applied Deep Learning (Supplementary)

[Paper Review] End-to-End 0:40:18

[Paper Review] End-to-End Object Detection with Transformers (DETR)

Комментарии
Автор

bro you're insanely good to explain complex stuff

vincenzo
Автор

17:50 for anyone asking why is that, that's the Generalized Intersection over Union Loss (LGIoU) that is defined as 1 - GIoU.
The GIoU range is [-1, 1]; when GIoU = 1 it means that there is a perfect overlapping (and then LGIoU = 0), meanwhile when GIoU goes to -1 it means that boxes not only do not overlap but their centers are far from each other too (and so LGIoU goes to 2).
So, in the end, best case scenario is LGIoU equal to 0, worst is equal to 2.
So for this reason LambdaIoU is positive (actually it is equal to 2, as found in the Appendix of the paper).

ghostcleanseIRL
Автор

Really nice video! Thank you for doing this! For the IoU part in the loss, I agree that the coefficient must be negative to since better alignment results in higher IoU, so the hyperparameter lamda_iou must be negative.

supang
Автор

I think the IoU loss defined in the appendix actually makes sense, it has 1 - at the beginning so it goes down also

briannlongzhao
Автор

It really helps me a lot!! Especially the match part! thanks!

leoliu
Автор

hey, man. The iouloss part I think it’s a different iouloss function and the paper didn’t show you. In that function when the iou gets bigger the result gets smaller, that make sense.

leoliu
Автор

Thanks for the great explanation! Why doesn’t the author compare results with YOLOv5 or any other single stage models?

ameynaik
Автор

It will be awesome if you also cover the how to use this coding part .

YourAI_agent