Object detection with Yolov8 | Data issues vs model performance

🎬 Timestamps ⏱️
0:00 Intro
3:15 Object detector
3:37 Data
4:44 Results
12:53 Outro
🌍 Community 👥
Join our Discord server: / discord
Support me on Patreon: / computervisionengineer
#computervision #python #objectdetection #yolov8

[@ComputerVisionEngineer]

Did you enjoy this video? Try my premium courses! 😃🙌😊 ● End-To-End Computer Vision: Build and Deploy a Video Summarization API bit.ly/3tyQX0M ● Hands-On Computer Vision in the Cloud: Building an AWS-based Real Time Number Plate Recognition System bit.ly/3RXrE1Y ● Machine Learning Entrepreneur: How to start your entrepreneurial journey as a freelancer and content creator bit.ly/4bFLeaC All my premium courses are available to the Computer Vision Experts in my Patreon. 😉 www.patreon.com/ComputerVisionEngineer

[@karimmerchaoui9736]

Grea video! Could you please create a video discussing the optimal setup for training a YOLO model with a custom dataset (quality, dimension of the images...) ?

[@ComputerVisionEngineer]

Sure, I will try to. 😃🙌

[@shubhamkumbhar6023]

Hi I had a question about incremental learning. Suppose I train a yolo model on 5000 images, but I get 500 new images every week (the object classes are the same 21). Is it possible to incrementally train the model?

[@andrewkarp]

It would be really helpful if you made a video on tactics for retraining with new data pre trained models and adding new classes. I’m encountering catastrophic forgetting and trying to figure it out.

[@gamingabd]

Great ❤❤

[@Xamy-]

Ive been training a segmentation model using 700 images currently @200 epoch, the loss function has a really good downwards curve even at the end but MAP plataeus, not sure what my next steps would be to investigate whats causing this (ideally would like to know the loss of each image in validation set). Do you have any advice you feel you can give me?

[@ComputerVisionEngineer]

Hi, the loss function has a really good downwards curve both in the training set and validation set?

[@Xamy-]

@@ComputerVisionEngineerThe training loss curve on the seg model is good but val-seg loss and val-class loss drop rapidly within a few epochs then plataeu (val-box loss still curves). Validation dataset is only 60 images. I will be adding an additional 600 images to my training dataset by next weekend (still annotating). Also the model is aimed at segmenting text, logos, speech bubbles and signatures, trying to see if its viable to use it to inpaint images at scale in finetunes

[@ComputerVisionEngineer]

@@Xamy- More images in the validation set may help you as well. About the mAP, what value does it plateau at?

[@Xamy-]

@@ComputerVisionEngineerThe mAP plateaus at 0.65, the relevant png charts can be seen by visiting imgur and adding the following (YT is blocking direct link) : /a/QqQhM3Z

[@Xamy-]

@@ComputerVisionEngineerThe mAP plataues at 0.65, "QqQhM3Z" is the album name on imgur

[@mohd.abdulghani6172]

hai, this is wonderful video, it will be helpfull if you made a video how to deploy yolov8 instance segmentation on android using using TFLite model, thank you

[@ComputerVisionEngineer]

Sure, I will try to. 🙌