How YOLO Object Detection Works

Here we introduce YOLO (You Only Look Once), a powerful object detection framework capable of real-time detection using a simple yet effective strategy.
Timestamps
--------------------
Introduction 00:00
DPM and R-CNN 01:35
YOLO algorithm scheme 02:50
Architecture 05:10
Target outputs 05:50
Non-max suppression 10:10
Loss function 10:45
Limitations 13:45
Summary 16:33

[@blaine_stl]

Most thorough explanation I’ve come across

[@sigmadews5403]

This is the best explanation of YOLO I found. Thanks!

[@deepbean]

Glad you found it helpful!

[@You_Only_LiveOnce]

This is the best explaination I found so far! From india 🇮🇳

[@user-lc7mu9bn3b]

so, what yolo does is basically what I have been doing for these captchas for many many years? I love the video

[@faithselby6015]

This was really clear and precise! thanks :)

[@nursah8815]

Simple , clear and excallent ! Thanks for the explanation.

[@matinmrv4213]

This is the best explanation on YOLO! Thank you very much.

[@mohsinjunaid8454]

very helpful and easy to understand Thanks

[@aresqard]

Cool!!!

[@shehz351]

Very nicely presented

[@shiva2874]

Thank you so much

[@hidayat1999]

The class probability map is only used to do loss calculations?

[@erfanasgari21]

About Limitations at 13:55 , first bullet point, didn't you say that the model predicts an one-hot class per bounding box? doesn't that mean that each grid cell can actually contain up to B classes? thanks for the amazing explanation by the way

[@Adrianfinantyo]

Hi! Your explanation was dope! Mind dropping the source or reference for that model accuracy-speed comparison table?

[@deepbean]

Appreciate your comment! The full table can be found in the original YOLO paper (arxiv.org/abs/1506.02640).

[@jensYSsck]

Does this also work for YOLOv8? because YOLOv8 is different from other versions that use free anchor detection. Thank You

[@Manisha-bj7ug]

4:06 why we take square root of w, h?

[@anynamecanbeuse]

Why don't use cross-entropy for the class loss?

[@deepbean]

Good question...some later versions (YOLOv3 onwards) use binary cross entropy to enable multi-label classification. Not sure why mean squared error was chosen for YOLOv1.

[@Stopinvadingmyhardware]

not, I look as much as I want.

[@user-kv8cd8yf2n]

Do I understand correctly that in NOOBJ part ground truth Ci is always zero and in OBJ part ground truth Ci is in [0,1]? kzbin.info/www/bejne/qafRamCui5yqoM0si=pA0UvT1dzNKjNEqv&t=766

[@JorgeRojas-ru8yb]

On Ground truth slide seem to be a mistake. 8:06 Check x = (16-10)/10 = 0.6 should be and similarly y = (44-4*10)/10 = 0.4

[@shehz351]

It's a modulus operator

[@testg4andmmm129]

4:10 are u robot? Click sll the boxes that contains bicyle