@Siraj Raval, can you comment or make a video on how YOLO is trained? Are the two parts trained on different networks and then combined? Or are they all trained in one go? More info would be appreciated.
Hey my name is naazim I have made this video on detecting actions in basketball match with Yolo, tensorflow etc Pls check it out if you are interested in this topic th-cam.com/video/0X6yTkXn-qQ/w-d-xo.html
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.
Joss Whittle is right, and Siraj comment startled me as well first time I watched. The derivative always points to the direction of the (possibly local) maximum. The gradient definition used in the context of backprop is not different. Even though in HOG it does not represent an error to be minimized, the property still holds.
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!
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.
Hi Siraj, just another killer tuto !!! Let me just add that windows users (like me by the way) might have difficulties to install darkflow. They can encounter a cl.exe exit code 2. To get around that you have to use the pip install . within the cross compiler x86_64 command prompt. To do that you just use the Windows key, followed by ctrl-tab and then type v on the keyboard. This should lead you to the Visual Studio command prompts list. Choose the right one and then go to the cloned darkflow dir to issue the pip command. Keep up the great work Man !!!
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!
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.
Oh, look, apparently now I have to binge-watch all the videos of this new channel that I just discovered. Honestly, at this point amazingly good channels like yours amount to a chore.
17:27. As a C programmer, I'm kinda offended, but at the same time I think the best language for machine learning is python and the best framework is tensorflow.
The Magic V, do you want to have a tutorial on Google Speech API, i.e., convert your speech into text! Watch this: th-cam.com/video/jc_-AIYvfKs/w-d-xo.html
i appreciate his videos, i am a programmer and i am thinking of staying aside anything that will destroy jobs. i guess i am just hating. if you think, the car is the most useful invention, and i am starting to think computers are not the answer to a "better world" :(. i am depressed.
I am in CS too ! :) Just like a knife, deep learning can be used for wrong or good things depending on whose hands it is in ! I think our ethics should be questioned instead, to make sure we understand the impact what we’re creating. Cheer up ! Personally, I am excited for machine learning, what a time to be alive! :))
Watching while eating breakfast in Saigon Vietnam....you are amazing...thank you for distilling the steps for how to configure and set up...going to have a lot of fun running g through your example.... :)
YOLO does seem to be a pretty good, some researchers I've chatted to are making it work for pothole detection: github.com/sekilab/RoadCrackDetector Saves anyone having to report a pothole again (in theory). It's an interesting time to be alive.
Good additional confirmation, but I think a distributed used of the anonymised accelerometers in phones is probably more effective. www.boston.gov/departments/new-urban-mechanics/street-bump
"we present a new cascading structure added SVM stages which employ the confidence values of multiple preceding Adaboost stages as input". ... also, just googling "viola and jones", wikipedia: en.wikipedia.org/wiki/Viola%E2%80%93Jones_object_detection_framework
"gradient - they just using this word for this" lol !!!! you did my day man ;) It is you guys in neural networks, who probably JUST use this word for something else. Gradient is the mathematical value that shows whether the function is rising or decreasing and how fast. And it is correctly used for those purposes. But anyway - cool video!
😂😂 wasted 22 min.....all u did was reading the lines and in the end u said follow the link in the discription👏👏 if that was the case u could have rounded up the video in 2 min 😤😤😤
Bro learning mein ego mat la, this video was useful for lots of folks including me. This video has now given me a direction as to which research papers to start with.
Gradient from HOG and BackPropagation are the same thing... Is a mathematical value given where a function increase or decrease. So Gradient may be where things get darker (less light), or whether you get less error.
Ok, here it goes. This was your best video because you slowed it down, spoke clearly and tou were more methodical in your explanations. This is your base to move from now. Excellent job!. Now you mix in a little mathematical and theoretical explanation ans people will be applying to you for grad school. ...to wizardversity and beyond.
would be great to see this from scratch. For ex. I have a live video feed of a concert and I want to classify the guitar the person is using. How would one 1. Set up the environment - don't skip over anything that you think people may know. 2. Train it on a set of images of guitar types to 3. Test the model with some video feeds 4. Then test live
Great explanation Siraj. This was a very quick way to get an overview in object detection algorithms. Gotta start a project with Yolo, because hey Yolo.
Isn't there a rather obvious optimization if you look at movies? Bounding boxes for the same object in the next frame are bound to overlap mostly with the bounding boxes in the previous frame, so we can pre-boost the confidence score for the small 13x13 grid boxes once a class is detected. And any score that is going to obscure that, will have to be higher than say half the original score. This way we can also quite easily follow objects and get better bounding boxes if they are partially obscured. Also by tracking their bounding box sizes we can estimate a 3D transformation and by averaging a 3D transformation of all objects detected get a idea of whether or not the camera is moving... So we can detect if a car is coming at us or driving of, or that the camera is moving towards a stationary object.. That would be great for self-driving tesla's I feel.
Man, this video was so great..and the yolo seems to be very good wth detection...I liked to be aware of it , thanks for sharing. Cheers and keep going. Brazil is watching..
I literally just sat down to do an assignment on this. Siraj, your timing is impeccable
thanks!
@Siraj Raval, can you comment or make a video on how YOLO is trained? Are the two parts trained on different networks and then combined? Or are they all trained in one go? More info would be appreciated.
I just liked this comment to bring the total to 69 :D
Hfish21 please can you tell me how did u do all this work... Because its my project work.. It need it at any cost please
Hey my name is naazim I have made this video on detecting actions in basketball match with Yolo, tensorflow etc
Pls check it out if you are interested in this topic
th-cam.com/video/0X6yTkXn-qQ/w-d-xo.html
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.
yeah I was surprised that Siraj didn't know that this was identical to a gradient.
I think he meant the gradients don’t have the same function as they do in backprop, i.e. representing an error value
So pretty much like a vector in physics.
Joss Whittle is right, and Siraj comment startled me as well first time I watched. The derivative always points to the direction of the (possibly local) maximum. The gradient definition used in the context of backprop is not different. Even though in HOG it does not represent an error to be minimized, the property still holds.
Yes one is gradient as in describing a slope, the other is gradient as in color. I think thats what he means by different :)
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!
You sir, are the reason my company is headed into softwsee development, coding, and programming. This video is worth more than gold.
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.
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.
Bro you might not know this...but you're pretty good at this TH-cam thing lol. Thanks man you're the best
The secret is use deeplearning to improve the video
Watch me man!
th-cam.com/video/jc_-AIYvfKs/w-d-xo.html
Thanks George lots of practice
teaching is the best way to learn
Hi Siraj, just another killer tuto !!! Let me just add that windows users (like me by the way) might have difficulties to install darkflow. They can encounter a cl.exe exit code 2. To get around that you have to use the pip install . within the cross compiler x86_64 command prompt. To do that you just use the Windows key, followed by ctrl-tab and then type v on the keyboard. This should lead you to the Visual Studio command prompts list. Choose the right one and then go to the cloned darkflow dir to issue the pip command. Keep up the great work Man !!!
Gotta send a link of this to my ex-wife! Maybe she can finally detect that I am a person.
Way to much info to much but it's good your venting.
Never marry a lizard person
haha wow thats real af
#LIZARD PERSON REALLY?/@#
Lol, i wish in future it can detect and read mind
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!
Its been five years. How about now?
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.
thanks Carl noted
Waoh, it's like passing from electrical signals to file processing. This is a technology miracle.
CNN works this time
1- Computation
2- Large Amount of Image available
I've seen YOLO before about a year or two ago it seems like it got better even since then. Good to see them still improving their product.
Man! You are amazing. your kind of presentation makes me stay completely focused!
I love this video!
Never expect any but the best from Siraj.
It seems that there's a faster algorithm called ssd multibox object detection, even works somewhat fast in android
yes.. ssd is faster than the yolo. and better suit for real-time applications.
@@kevaldholu7366 well yes. But the yolo is more favored because its simplicity than the latter.
This is crazy I graduated back in 2012 and things have changed a lot.
Object detection made easy
plz tell me how to implement this on my Windows PC ..plz tell me some way out for this bro.. ....
Buy a MAC
Bro this isn't a valid solution..
The dark net has also windows version.. but i haven't know complete knowledge to set environment on Windows
Use VoTT and CNTK docs.microsoft.com/en-us/cognitive-toolkit/object-detection-using-faster-r-cnn
wahh !!!!! what an explanation man ??? mind blown for 30 mins straight
I'm a genius now, thx
@Xingming Pinyin this will make you genius, xigishihiwifisidirixieitiyiuiiy
Thanks for your work it is the first time i find proper and clear explanations about how to interpreter the network output.
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.
Oh, look, apparently now I have to binge-watch all the videos of this new channel that I just discovered. Honestly, at this point amazingly good channels like yours amount to a chore.
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
17:27. As a C programmer, I'm kinda offended, but at the same time I think the best language for machine learning is python and the best framework is tensorflow.
I was looking for this just a few days ago and was a great coincidence that you decided to upload this video , thanks!!
i love how siraj's videos are understandable until the last quarter or so and then it's a freaking downhill
I was about to do my assignment on YOLO on Deep Learning Specialization by Andrew Ng, and this pops out right on time!
@siraj, at 0:50; And we are going to build our own model as well....". But, when did we build our own model???
THE BEST DESCRIBTION I SAW. REALY NDERSTANDABLE
Siraj, can you do a video showing how to install YOLO, please? Thank you so much
Dude your channel is pure gold
0:41 i'm still searching for the train!
Thanks for your work it is the first time I find proper and clear explanations about how to interpret the network output!
I think it's developed by Joseph Redmon ...... YOLO i've seen his TED talk. and also he made it as open source.
It is such a beautiful initiative taken by you to teach the globe about the threshold technologies. Keep the good work up.
Yeeeee I waited for so long for yolo
The Magic V, do you want to have a tutorial on Google Speech API, i.e., convert your speech into text!
Watch this:
th-cam.com/video/jc_-AIYvfKs/w-d-xo.html
That was an excellent description of a topic that has been confusing the heck out of me for many hours. Thank you!
From a guy who defined the concept of a "logic door"...
I LOVE THE FUTURE!!! YOU R A ROCKSTAR Siraj!!
wow it detects MP4 recorded files and in "Real Time" cooooool
Any ideas why image jpgs work fine and mp4s don't on my mac?? thanks!
Thanks so much for this video Siraj, I really enjoy that it doesn't have as many cuts as usual
"we are going to build"?. i think you used someone else's code. 20 min of explanation and 2 of demonstration ::thinking::
L Lawliet yeah it was done by someone else
ok you can do 20 min of demo and 2 min explanation, but you wont. So why, because never believe in appreciation
i appreciate his videos, i am a programmer and i am thinking of staying aside anything that will destroy jobs. i guess i am just hating. if you think, the car is the most useful invention, and i am starting to think computers are not the answer to a "better world" :(. i am depressed.
I am in CS too ! :) Just like a knife, deep learning can be used for wrong or good things depending on whose hands it is in ! I think our ethics should be questioned instead, to make sure we understand the impact what we’re creating. Cheer up ! Personally, I am excited for machine learning, what a time to be alive! :))
Cynthia Habonimana will all laugh when theses fuckers of ai learn to code
Watching while eating breakfast in Saigon Vietnam....you are amazing...thank you for distilling the steps for how to configure and set up...going to have a lot of fun running g through your example.... :)
nice video, plz make more
you are such an awesome human being Siraj
Heads up, version 3 is just out: pjreddie.com/darknet/yolo/
Paper: pjreddie.com/media/files/papers/YOLOv3.pdf
Thank you sir!! Your pronunciation is very well ,amazing ! I understand without subtitles thank you this informative video and your expression
Anyone got any opinions/warnings regarding YOLOv3? About to start a project and dont wanna make my life more difficult than it already is
Yolo V3 is really simple. I have some experience with it :)
Man, this is one of the best explanation for begginers I ever see !!!!! Very good. Do y have any ML course? I will pay for it.
Oh how I miss 2018 machine learning.
hi siraj,
when I doing the YOLO I encountered: AssertionError: Over-read tiny-yolo.weights.
the env is win7 and python 3.6.3.
Great Tutorial will train to recognize Students and Faculty and objects like weapons.
I thought SSD, faster R-CNN using ResNet, and R-FCN were all more accurate than YOLOv2
you are right yolo is fast but not accurate as other architectures
will look into SSD
I luv your hands-free scrolling in this video
imagine doing this but with capsule !! new project idea !!
sprinkle capsule on everything lol
Siraj, even though i don't do anything AI related, I always watch your videos just in case I get started. I've learned so much
Hey siraj nice work out there
I am trying to start AI can you give me some recommendations about the content and there order to learn.
Thank you.
he has a playlist in his youtube page where he shows how it work
See the video quick questions with siraj raval on this channel itself. You will find your answer!
my playlists
Outstanding explanation and I appreciate the way you presented your project.
Keep illustration
You're just reading off from a github page that's all and for the implementation part you just flew through it......!!
Still helps.
The important part is it helps.
The reason you are here is also the same.😂
Thanks for the great explanation. I now understand the significance of YOLO.
YOLO does seem to be a pretty good, some researchers I've chatted to are making it work for pothole detection: github.com/sekilab/RoadCrackDetector
Saves anyone having to report a pothole again (in theory).
It's an interesting time to be alive.
great link!
Good additional confirmation, but I think a distributed used of the anonymised accelerometers in phones is probably more effective. www.boston.gov/departments/new-urban-mechanics/street-bump
Reckless Roges why not both? It's always good to crack the same problem in many ways.
Needless to say, subscribe at once, a very clear and useful presentation.
I made this few months back for my college techfest. checkout this ,it is the one that inspired me> github.com/oarriaga/face_classification
Very good work, I will fork for sure :D
very cool
Will it assume my gender??? omg im getting triggered...
I am glad to hear that :)!
Mehn!! you are a great teacher I wish I could subscribe a thousand times. Thank you for this♥️
Can you show and explain SSD too? :3 I've read it should be faster then YOLO :)
Nope, it's not. YOLOv2 is the fastest object detection out there. Check their comparison here (pjreddie.com/darknet/yolo/ )
will consider ssd
Did you?
This is Brilliant. I'm actually gonna play with it. Thanks Siraj!
violo jones uses svm? omg can't you google stuff before you talk?? viola jones are famous for combining cascades of boosted classifiers...
the improved version uses SVM link.springer.com/chapter/10.1007/978-3-642-22822-3_7
"we present a new cascading structure added SVM stages which
employ the confidence values of multiple preceding Adaboost stages as
input". ... also, just googling "viola and jones", wikipedia: en.wikipedia.org/wiki/Viola%E2%80%93Jones_object_detection_framework
Hey dude thanks so much for your lengthy explanations and your enthusiasm when you make your videos. It really helps !
stop objectifying dogs siraj /jk
lmao
lol
"gradient - they just using this word for this"
lol !!!! you did my day man ;)
It is you guys in neural networks, who probably JUST use this word for something else.
Gradient is the mathematical value that shows whether the function is rising or decreasing and how fast. And it is correctly used for those purposes.
But anyway - cool video!
his movements are irretating me :(
Awesome video. Time to combine YOLO with my 3d printed Arduino powered robotic arm.
Congrats, you know how to copy-paste. But just barely.
Wow - impressive! The technology - but even more the way your way to explain it!!
Could it recognize person in hijab?
You are a GEM!
This is the best explanation of object detection I've watched. Great work Siraj
😂😂 wasted 22 min.....all u did was reading the lines and in the end u said follow the link in the discription👏👏 if that was the case u could have rounded up the video in 2 min 😤😤😤
he spent that time to explain to you in detail what was happening you ingrate
he explained it really well but i agree
Wow! People wanna code without knowing the logic behind the code. What has the world come to? 🙈
Bro learning mein ego mat la, this video was useful for lots of folks including me. This video has now given me a direction as to which research papers to start with.
Siraj! Thank you so much! When you explain step by step like this I can undestand everything! Love this video!
i will literally pay for enjoy this content in spanish
I Love all your videos. You are precise, fast, make mountainous task so simple to deal with... Thank.
Gradient from HOG and BackPropagation are the same thing... Is a mathematical value given where a function increase or decrease. So Gradient may be where things get darker (less light), or whether you get less error.
great video, thorough explanation
Really amazing explanation!
I did my phd on image segmentation around the turn of the century, and i remember waiting hours to process one image. How far we've come.
This is really awesome. You explain it in such a clear and simple way.Thank You!!!!.
Ok, here it goes.
This was your best video because you slowed it down, spoke clearly and tou were more methodical in your explanations. This is your base to move from now. Excellent job!. Now you mix in a little mathematical and theoretical explanation ans people will be applying to you for grad school. ...to wizardversity and beyond.
Best video yet! I like this lecturer-y style much more, keep it up!
Amazing video.. thank you very much!!!
would be great to see this from scratch. For ex. I have a live video feed of a concert and I want to classify the guitar the person is using. How would one
1. Set up the environment - don't skip over anything that you think people may know.
2. Train it on a set of images of guitar types to
3. Test the model with some video feeds
4. Then test live
Windows environment ?
Great explanation Siraj. This was a very quick way to get an overview in object detection algorithms. Gotta start a project with Yolo, because hey Yolo.
Isn't there a rather obvious optimization if you look at movies? Bounding boxes for the same object in the next frame are bound to overlap mostly with the bounding boxes in the previous frame, so we can pre-boost the confidence score for the small 13x13 grid boxes once a class is detected. And any score that is going to obscure that, will have to be higher than say half the original score. This way we can also quite easily follow objects and get better bounding boxes if they are partially obscured. Also by tracking their bounding box sizes we can estimate a 3D transformation and by averaging a 3D transformation of all objects detected get a idea of whether or not the camera is moving... So we can detect if a car is coming at us or driving of, or that the camera is moving towards a stationary object.. That would be great for self-driving tesla's I feel.
Wow you just made my day!
Now, I think i’v created the best bot for a game.
bless you
so informative yet so easy to understand
10/10 for this. I'd never heard of YOLO, and this is a really great introduction.
Video quality has gone way up! Nice job Siraj!
thanks john
Very informative! Thanks siraj 😊
Thank you for the great video!! Really helpful with my projects!!
Man, this video was so great..and the yolo seems to be very good wth detection...I liked to be aware of it , thanks for sharing. Cheers and keep going. Brazil is watching..