YOLO Object Detection (TensorFlow tutorial)

You Only Look Once - this object detection algorithm is currently the state of the art, outperforming R-CNN and it's variants. I'll go into some different object detection algorithm improvements over the years, then dive into YOLO theory and a programmatic implementation using Tensorflow!

Code for this video:

Please Subscribe! And like. And comment. That's what keeps me going.

Want more inspiration & education? Follow me:

More learning resources:

Join us in the Wizards Slack channel:

And please support me on Patreon:
Signup for my newsletter for exciting updates in the field of AI:

[yetBnAmd]

I literally just sat down to do an assignment on this. Siraj, your timing is impeccable

[Loopyengineeringco]

TBH, I only clicked this because it said YOLO. Now my brain is exploding.
But joking aside, you're a great explainer and this is all starting to make sense. Thanks for the video!

[medmed]

Man! You are amazing. your kind of presentation makes me stay completely focused!

[schulca]

These videos are great! also a lot easier to focus on when there aren’t memes popping up all the time. I enjoy the lecture style.

[JossWhittle]

At 4:10, HOG does actually mean Gradient in the same way as backprop does. An image is just a discrete representation of a continuous 2D signal, the gradient of the continuous signal at a point can be approximated from the discrete representation by taking the finite difference between neighbouring pixels.

[georgebockari]

Bro you might not know this...but you're pretty good at this Youtube thing lol. Thanks man you're the best

[RatherBeCancelledThanHandled]

I thank God, that I started studying programming/math, so much fun and so fascinating to be able to take part in such cool technological advancements.

[yashchandraverma]

CNN works this time
1- Computation
2- Large Amount of Image available

[planktonfun]

It seems that there's a faster algorithm called ssd multibox object detection, even works somewhat fast in android

[jbuist]

That was an excellent description of a topic that has been confusing the heck out of me for many hours. Thank you!

[Lavimoe]

The whole video is very thorough and comprehensive, which makes such intimidating subject a no-brainer for the beginners. Not sure how I will use YOLO in my future projects, but I really learned a lot from this video!

[Lunsterful]

Gotta send a link of this to my ex-wife! Maybe she can finally detect that I am a person.

[intrvrt]

Object detection made easy <3 Thanks siraj 😁

[josephfoltz]

You sir, are the reason my company is headed into softwsee development, coding, and programming. This video is worth more than gold.

[oliviersaint-jean]

For videos, I think the algorithms should take the time dimension into account, (ie. increasing the probability of an object detected in one frame to be there again in the next frame) to decrease computation cost.

[jazzpote]

Your videos are so amazing. You cover all the fields of CS practically, with a state of the art approach.
So helpful, keep it up

[exratt]

Hi Siraj,

thanks for your video. I never heard of the YOLO detector before and find this approach very interesting, as I'm used to the good old brute force method of detecting objects. I have a few remarks concerning the two mentioned pre-deep-learning algorithms.

Regarding the Viola-Jones detector: The features are hand-coded (Haar-like features, which are basically the gray-scale value difference of neighboring rectangular regions), but the locations are not selected by the researchers themselves, as suggested by your video. Instead, they are selected by the training algorithm. They did not use a support vector machine for classification, but a cascade of simple classifiers, which were trained using AdaBoost. Maybe you confused it with the HOG approach.

What made the Viola-Jones detector so efficient was the features and cascade. The features could be computed very efficiently using an integral image (only three additions to compute the sum of gray-scale values over any axis-aligned rectangular region). The cascade was trained such that image windows which did not contain a face would be discarded very quickly, so only very few windows needed to compute all the features and go through all cascades.

The image on your slides is also a bit misleading. It mentions local binary patterns, which is another feature extraction method. The image shows face recognition, in this very case to find out whether a face belongs to the person it pretends to be.

The Dalal-Triggs detector uses histograms of oriented gradients, as you mention. They build histograms over each cell, so it does not only contain the strongest gradient direction of all the pixels in a cell.

[yannickmolinghen]

Thanks for your work it is the first time I find proper and clear explanations about how to interpret the network output!

[MrZouzan]

I was looking for this just a few days ago and was a great coincidence that you decided to upload this video, thanks!!

[mirandaclace]

Anyone got any opinions/warnings regarding YOLOv3? About to start a project and dont wanna make my life more difficult than it already is