Build computer vision software - Q&A, problems and solutions
ฝัง
- เผยแพร่เมื่อ 15 เม.ย. 2024
- I selected some computer vision questions I received and answered them in the video
🌐 pysource.com/
➤ Courses:
✅Full Computer Vision course:pysource.com/object-detection...
➤ Follow us on:
LinkedIn:
/ pysource
➤ For business inquiries:
📊 pysource.com/contact - วิทยาศาสตร์และเทคโนโลยี
Thanks a lot sir, I'm currently working on the project which is about preventing students from cheating. I'm already broke the problems down into some areas:
1. Eye gazing to check if the students are looking outside.
2. Eye blinking the keep track of the abnormal behaviors.
3. Mouth opening to check if they are speaking or not.
4. Object detection to check whether there are many people on the camera and whether they use untrustworthy devices.
5. Noise detection to check whether a abnormal sound.
6. Head pose estimation to check whether they looking outside, maybe I should combine this with the eye gazing, if there's no eyes, there will be abnormal behaviors.
7. Face spoofing to check whether the face on the screen is real face or not.
I don't really know which parts should I put more into the project or should I remove something, because it's quite a challenge for me to get into the intern place in the project. I'm already done with the first three and planning to finish all the stuffs. Pls let me know your opinion about the idea and other question is how many features are good for me to show my effort to the project manager?
Again, thank you sir and have a nice day!
Thanks a lot, Sergio!
I didn't expect to be selected for the video to be honest, otherwise I could have given more details.
But you guessed the setup exactly right.
We are taking a picture from the top and then I need to analyse the picture of each product in detail (at some specific x/y coordinates) and based on that judge if it is OK or NOK.
My camera currently is at a height of around 30cm from the product and we take 4K videos/pictures.
The specific area within our part - that we need to analyse - is approx. 1,5cm x 1,5cm.
1. Object detection
2. Object classification
Would also have been my approach. Unfortunately it is a very specific part in automotive industry and there isn't much material that can be used for training.
Thanks and best regards,
Dali
Your approach Object detection first, and Object Classification later is good. Some tips:
- Controlled lightening is very important and this should be considered as an important part of the design
- Industrial camera with high resolution and high shutter speed if objects are moving fast
- after you detect the area, zoom in and cut the region you need to analyze with object classification
- training should be done on the real samples and if there is not enough material you can use augmentation techniques to increase the data).
@@pysource-com thanks
1. lighting - yes, we got that covered
2. the belt is very slow, so that shouldn't be a problem
3. after detection, zoom in and cut the region - ok, good point. This will be the tricky part.
4. training data - yeah, thanks
for zooming in and cutting the region I am thinking I might get into the topic with an angle, since my object is quite big and the areas of interest are spread around the object. If I want to work just with 1 picture than I might have to adjust my picture to make it a rectangle again (as you also explained in the video)...
Thanks!
Great Dali, do those steps and you'll be on the right way to successfully complete the project.
Bardzo edukacyjny film 😊 Piaski
Dzieki ;)
Amazing video, I'm waiting for new episodes.
I'm working on a project to detect counterfeit products, for example airpods boxes, where we have a high level of quality in counterfeiting, making it difficult for a human to analyze without having an original box to compare.
But I can see differences in the font used, the spacing between lines and the product photo has slightly less contrast than an original.
I intend to do classification using YOLO, but the dataset has few images and they will all be very similar, leading me to try to use Siamese neural networks, few shot learning, to find the most similar one and classify it.
My question, is there some approach I'm missing? thanks!
Thanks for the feedback.
The project you're working on it's a bit complex as you'll be identifying details that require very high precision, so it's not easy to have reliable results.
You're approach is a good starting point to make a study, get benchmarks and see how it performs, then you can evaluate the next steps of you projects.
best of luck with your projet
i wanted to thank you for all the efforts you have put in all of your videos ❤❤
i really have learned a lot from your videos especially neural networks series
i am a student in the faculty of computer engineering and i am currently in my bachelor thesis semester and it is about vehicle counting and classification
now there are ton of models and detection algorithms can you recommend the best and easiest to work with models or algorithms out there for example there are 8 or 9 versions of yolo but i do not know what would be the most stable or best one out there
thank you in advance
Thank you for the feedback, I'm glad the neural network series were useful to you ;)
About your question, for vehicle tracking and classification, YOLOv8 by Ultralytics is a good choice.
@@pysource-com THANK YOU SO MUCH FOR YOUR REPLY
ive been reading about YOLOv9 lately and the technologies used like PGI or the new architecture GELAN are outperforming YOLO v8
so i wanted ask if you have tried YOLOv9 and if yes do you find it easy to work with and stable