How to Capture and Label Training Data to Improve Object Detection Model Accuracy

[@EdjeElectronics]

Unfortunately, Cleo passed away unexpectedly several months ago. She was a wonderful cat, and she will be dearly missed. May her memory live on forever through this KZbin video.

[@charuwaka1]

😭😭😭😭

[@calvjohn8165]

😢

[@ciscomike82]

😪

[@deplorablesecuritydevices]

Sorry to hear about Cleo.

[@AbhinavMishraA-wc3ky]

😭

[@nickfisher6813]

this has channel has been a huge help for my college projects, thanks for all your uploads, literally way to helpful

[@ivandimitrov2401]

Amazing tutorial! One of the most useful videos I've ever watched. Keep up the good work! Thanks a lot, Edje!

[@kevinsmithwoo]

Please, keep creating content. Awesome job!

[@orkunvemosi]

sorry about your cat. you're an awesome human being for teaching us this stuff

[@mage2754]

Thank you so much. Been looking for a concise explanation on this

[@TomasAlmeida-CC]

Thank you, that was really useful, and your cat is so cute!

[@syazwansharani8907]

thanks! I learnt a lot from this. Cheers, mate :) and RIP Cleo ;(

[@piconano]

Thank you for your informative video. I'm starting on this journey and have a few questions before I go too deep down the rabbit hole. 1- If I had a video showing an object from all possible positions, is there a way to automate the training and let a script take picture instead of me? 2- Once this dataset has been captured, is it possible to transfer the dataset to an identical system without retraining? What about the improvements that each machine will come up with, is there a way to integrate new learned data and share them between two or more machines? I have more questions but for now, these are my stumbling blocks. Thanks again.

[@nurmanprihatna4274]

Awsome ... it very inspiring and gave me new insights about computer vision..

[@dcdales]

This is very helpful. Thanks so much, Edje!

[@te-te3gz]

In our project, we plan to use the YOLO model for real-time object detection, trained with the COCO dataset. Both the dataset and the class information will be stored locally on the Raspberry Pi. Our main concern is whether the Raspberry Pi has sufficient processing power to handle the real-time demands of running the YOLO model with this dataset, or if we would need to explore optimizations or additional hardware.

[@microwavecoffee]

When dealing with images not taken by your camera, would you suggest normalization of images dimensions before/after drawing bounding boxes, or to do it programmatically?

[@MrMintMonkey]

Gained a new subscriber, keep up the great work!

[@ANUJSINGH-bz7ci]

love u mate ❤ u help me alot doing things ,huge respect

[@EdjeElectronics]

Thanks man! I appreciate it. More videos coming in the next few months :)

[@cbara568]

Please, when is the next video coming out? I can't wait

[@christophgottert3110]

Thanks for the video! I wonder how to decide which resolution & aspect ratio to choose for the test data? E.g. if I'm about to use a model with 256x256 input tensor, does it even make sense to label an 800x600 image?

[@SA-oj3bo]

Hi, could you please make a video how to recognize sounds with a RP? Thanks in advance!

[@calvjohn8165]

This is an amazing content, I have tried the coins detection as well and it worked but how do you change the color of the bounded boxes?

[@EdjeElectronics]

Thanks! Check Line 209 of my TFLite_detection_webcam.py script - the "cv2.rectangle" call sets the color of the rectangle to (10,255,0) which is a light green. You can change that value to change the color. github.com/EdjeElectronics/TensorFlow-Lite-Object-Detection-on-Android-and-Raspberry-Pi/blob/master/TFLite_detection_webcam.py#L209

[@misticmuse-b4c]

Thank you so much

[@lontongtepungroti2777]

your video is amazing thank you very much :")

[@romellfudi]

Thats awesome, how many objects do you need per class?

[@EdjeElectronics]

I recommend using at least 50 images per class to start with 👍

[@danutaishleash7334]

Is it possible to use the object detection feature to detect and identify small components on PCB boards. like marking on resistors and IC.s chips?

[@EdjeElectronics]

Yes, but you would have to use something like Optical Character Recognition (OCR) to read the markings on the resistors and ICs. That would be a fun project!

[@rubialugattimoreira1978]

Perfect!

[@briannyirenda9951]

Thanks for this awesome video am really learning ... but I have a question I am creating a social media app and decided to use tensor flow lite for content moderation is it possible to prohibit images that have been uploaded by users if they are in appropriate. ?

[@ibersonsilva9413]

Good video! I just have one question: How can I detect places instead of objects?

[@patriciowallace446]

Hi, very interesting, There is a way to presize the size of an object if I fix the camera in a place and an object in the same perspective and distance, like a machine making parts, just to detect parts not good well done by the machine. There is a way ?

[@EdjeElectronics]

I see what you mean! Unfortunately, LabelImg is fairly basic and I don't think there's a way to presize the bounding box. You might want to look at Label Studio - it's free and offers more flexibility. labelstud.io/

[@faraway27]

Thank you very much for your tutorials, but I am still having an open question: I want to create a shopping item detection model, which detects items in supermarkets. Is it possible to label these objects always as „items“ (0, Background and 1 item), since the camera will only see background or shopping items? Or will it be difficult to train such a model on only one class?

[@EdjeElectronics]

Oh yeah! I remember your question from the other video. If you're detecting items on a checkout conveyor, I think it would work to just label everything as "item". You don't need to worry about having a separate class for background. Your labelmap would have just one entry, like this: item If you have enough images of objects on the conveyer belt and label all the objects as "item", I think it will be fairly accurate. I would start with just 100 - 200 images and see how it works. The only drawback is if you try the model on a new item on that it hasn't ever seen before, it might not recognize it.

[@faraway27]

Thank you for the answer. I thought so too, which is why I have already collected data. However, as a beginner, I did not realize how important the quantity as well as quality is, as shown here in the video. Now that I have seen this video I am excited to start the project properly.

[@sgschusswaffenlol9]

Hello- Thank you for this great tutorial. I am new to deep learning and image object detection. However I would like to detect a line or vector in my image. Are there any "Image Labeling" tools that help me to create my dataset,? Are there any tutorials online that show how to detect "line-like" objects? I would really appreciate it, if somebody can give me some tips.

[@RED-zs4cq]

labelstudio is also a good labeling tool

[@marc-antoinepelletier4661]

do you need to have all the objects in all the pictures or you could have 4 different folder for all diffrent type of objects?

[@GNPWIJAYARATHNA]

Hello, thanks for sharing, I want to know the image size or resolution, can you suggest any?

[@EdjeElectronics]

If you know what resolution your camera will run at in the application, then you should use that resolution. Otherwise, I generally recommend capturing 1280x720 images.

[@nipunagnp]

Thanks. Image Captures around 2-3 mb. My worry is this increase huge image capacity. And effects to collab’s processing time? . Your images are around 100 kb 🙄🤔😰

[@EdjeElectronics]

@@nipunagnp yep, if your images are 2-3MB, you’re probably using 1920x1080 resolution or higher. Try reducing it to 1280x720, or even 640x360.

[@khenpahilanga9596]

hello sir, thanks alot for this video, I would like to ask, is it possible to use object detection more specifically to detect disease on plant leaves, e.g. tomatoes or eggplants? I would highly appreciate it if you could provide tips if otherwise, to create a system for this kind of detection real time, im new to machine learning and these stuff :)

[@EdjeElectronics]

Hello! Here's a video that looks like it covers how to detect diseased plant leaves using a YOLOv5 model. When collecting the dataset, keep the tips from my video in mind. kzbin.info/www/bejne/o3fVn4V_i89ro5I

[@sohaibzafar6110]

your cat is beautifull MA SHA ALLAH

[@WAWASAMBHUWARA]

Really love yout video sir. I have a question. Can I use Faster RCNN on android? Because I try it but it doesn't work. Thanks

[@amberborn9236]

did anyone try this on mac recently? struggling to download. I assume I have to install python first?

[@mehow99_91]

Hey, I've Got a question. I was trying to run tensorflow object detection program an somehow have remote Access to this screen, not only from home but for example when I go to the nearby store and want to check if everything is safe iside my house. Could you help me with that?

[@arya83o292]

Sir, how to use feature extractor ssdlite mobilenetv3 small in tensorflow 2?

[@-0-__-0-]

are there no apps that does the annotation automatically?

[@HaiLe-hl1wj]

hello, how many images do you need per object for training it?

[@ajarivas72]

Around 2000

[@adityapatel-qp2ir]

Edje please help me. With ur batch size setting my model stops learning in initial phase with ^c . Without my intervention. Do u have any optimization or fix please help

[@07_firzaichlasulamalarians8]

hei why i use same format xml but my dataset get error in convert to csv but if im use ur dataset it can can u solve that?

[@JoshuaDowell-h8o]

Can this product export item and time into a excel spreadaheet

[@ThatGuyThatCommentedOn_A_Video]

Nice dog

[@mr.sunshine0226]

the cat want to escape 😼

[@daryljaysagabaen9562]

8:23