How much data you need to train a compute vision model?

Рет қаралды 4,283

Күн бұрын

Пікірлер: 47

@ComputerVisionEngineer 9 ай бұрын

Did you enjoy this video? Try my premium courses! 😃🙌😊 ● Hands-On Computer Vision in the Cloud: Building an AWS-based Real Time Number Plate Recognition System bit.ly/3RXrE1Y ● End-To-End Computer Vision: Build and Deploy a Video Summarization API bit.ly/3tyQX0M ● Computer Vision on Edge: Real Time Number Plate Recognition on an Edge Device bit.ly/4dYodA7 ● Machine Learning Entrepreneur: How to start your entrepreneurial journey as a freelancer and content creator bit.ly/4bFLeaC Learn to create AI-based prototypes in the Computer Vision School! www.computervision.school 😃🚀🎓

@uniqx-ai 9 ай бұрын

Can I know your upwork or freelancer? I have the project

@wijzeuil 4 күн бұрын

This is amazing! I woke up today knowing nothing at all about object detection, and now I’ve somewhat achieved my goal of detecting lighthouses in a live surfcam! I wonder what would happen if I increased the number of epochs as well. I’ve noticed that with my dataset of 20 training images, I get poor results with 20 epochs but achieve close to 60-70% recognition with 100 epochs. Even though the results indicate 99%+ precision, I don’t actually have 99% of the images labeled correctly. That’s why I really appreciate your visual verification approach. Numbers and graphs can be misleading for inexperienced users like me.

@tonymudau3005 9 ай бұрын

Thank you so much for this tutorial! Exactly what I was trying this weekend. Spent my time annotating 3000 coca cola images with 26 classes. My model is so bad right now. Something that I experienced was over fitting. Your ducks video is very simple but it is likely that if there was more complex background you would see over fitting on 4000 images. Another issue was that for some reason when you have many classes and maybe other classes have more images. When tracking you can experience a complete switch of labels.

@ComputerVisionEngineer 9 ай бұрын

Glad it is helpful! Yeah in my case it was only one class, so it was simpler. 🙌

@felipeyanez988 7 ай бұрын

Thanks, I just needed this, I was training +2000 images per dataset with very limited resources, now I will try with 500 images to improve performance. Saludos desde Chile!!

@ComputerVisionEngineer 7 ай бұрын

Glad it was helpful! Saludos! 😃🙌

@afjamo 7 ай бұрын

Wonderful! I was wondering why prediction didn't work with your alpaca tutorial video. No I completely understood. I tested my scripts with only 56 images and did not get any prediction. I will increase the number of images! Bravo! I appreciate your work!!

@ComputerVisionEngineer 7 ай бұрын

Glad the content is helpful! 😃🙌

@RAHUL-dt5xm 9 ай бұрын

keep it coming bro, your videos are informative and helping me learn new things.

@ComputerVisionEngineer 9 ай бұрын

Glad the content is helpful! 😃 I will keep it coming! 💪💪

@TheJAM_Sr 7 ай бұрын

I love your channel dude! Thank you for all the instructions and testing. I’ve been able to learn a ton and apply it to my own project. I do have a question about training yolo models: I have a model I’m training with 3 classes, 2 of the 3 classes train much higher than the third. Is there any recommendation you have to even out the training? Is it all dataset? Should I be making sure that there are more of 3rd class than the others? I appreciate you!

@ComputerVisionEngineer 7 ай бұрын

Hey thank you! Glad the content is helpful. Is your dataset balanced? (balanced = same amount of objects in each class)

@TheJAM_Sr 7 ай бұрын

@@ComputerVisionEngineer I believe so. I would say out of the 500 - 1000 annotated images there is less than 5% that don’t have all 3 classes present. But that is a good point, I’m going to double check that.

@GKEMRECAN 9 ай бұрын

Thank you for this tutorial

@ComputerVisionEngineer 9 ай бұрын

You are welcome. 🙂

@uniqx-ai 9 ай бұрын

This might be about the amount of data, what about the influence of the images themselves? What is the ideal situation? Should it be very varied? Or is it okay if they are similar, say the background is still the same but the objects are from different perspectives? I'm still confused about the images, sometimes they can't be found on the Internet and I have to take them myself (in the case of my project). Thank you in advance. Your videos have been very helpful.

@ComputerVisionEngineer 9 ай бұрын

I have other 'experiments' in mind to see how the performance is affected by different type of issues with the data. Analyzing what happens if the background changes or if it is always the same is a good idea, I could make a video about it! 🙌

@tankaccount2990 9 ай бұрын

In my experience if you keep the background the same the model starts detecting the background as the image to detect instead of your desired object

@yoshimochii 7 ай бұрын

great video! just wanna ask, how did you split your data into train and val? is it 70-30?

@ComputerVisionEngineer 7 ай бұрын

90-10 if I remember correctly

@aadiduggal9762 9 ай бұрын

Hello. I was trying to code your python sign language detector, but when I ran the code to collect the data of my hand symbols, I couldnt find anything in my data file. Please help

@fitox1234 9 ай бұрын

Could you create a video demonstrating how to develop facial recognition for tiredness and stress?

@ComputerVisionEngineer 9 ай бұрын

Do you mean an image classifier the input is a face and the output is tired / not tired?

@fitox1234 9 ай бұрын

@@ComputerVisionEngineer I have tried to do recognition of more complex emotions such as stress and fatigue on the face, but I get very similar results among them and with the neutral state. How would you start that project? 👀 Can you make a video about it?

@ComputerVisionEngineer 9 ай бұрын

@@fitox1234 oh I see, yes it is very challenging to tell those categories apart, I will try to do a video about it.

@MDAsadullahShibli 9 ай бұрын

Can u share any intermediate or expert label project u did on upwork ? What would be great help for us🥰

@ComputerVisionEngineer 9 ай бұрын

Sure, I will find some. 🙌

@F_F_CH 8 ай бұрын

Ciao Filipe can you do same thing with various model generation tools(YOLO, Scikit, Teachable... ), pls?

@ComputerVisionEngineer 7 ай бұрын

I will try to 🙌

@F_F_CH 7 ай бұрын

@@ComputerVisionEngineer Anyway, for clarification: Could you do a model accuracy test for various tolls like YOLO, Scikit, Teachable. I feel that they are made for various types of problems, and can you do add which toll is better to use for which problems.

@shubhamvashishth8289 9 ай бұрын

Bro could you please make a GAN tutorial playlist along with some projects ?

@ComputerVisionEngineer 9 ай бұрын

I will try to make some videos about gans. 🙌

@NathanCamillerii 7 ай бұрын

Would this also apply to keypoint detection using YoloV8 please?

@ComputerVisionEngineer 7 ай бұрын

I think it is likely to have similar results in a keypoints detection problem.

@adrian_maulana3404 9 ай бұрын

Thanks you for your tutorial,btw can you make tutorial yolo v9😁?

@ComputerVisionEngineer 9 ай бұрын

I will try to. 🙌

@adrian_maulana3404 9 ай бұрын

@@ComputerVisionEngineer ok thanks bro👍

@thomschery2800 9 ай бұрын

Let's say he wants it to determine which player won a duel in some game based on the uploaded video. Is this possible?

@ComputerVisionEngineer 9 ай бұрын

Not sure if I understand, what is the machine learning problem you are trying to solve?

@thomschery2800 9 ай бұрын

@@ComputerVisionEngineer Which player won the game. Would AI be able to give me such information?

@TheJAM_Sr 7 ай бұрын

What information is available on screen to determine that? You can train the model to detect health bars, score, ect

@thomschery2800 7 ай бұрын

@TheJAM_Sr OK, so what would be the most preferable tools?

@TheJAM_Sr 7 ай бұрын

@@thomschery2800 you trolling? Your questions are so vague. If you’re asking what annotation tools, it really depends on you. I like to run everything locally so label studio is my choice. There is plenty of tutorials here for beginners.