The blood components are unbalanced - we have predominantly more red blood cells compared to the other types. But the visual characteristics of the 3 types studied here (red blood cells, white blood cells and platelets) are sufficiently distinct so that the model is able to perform well despite the imbalance.
Thanks for sharing your project @maria_rodriguez! The post was really easy to follow.
It’s interesting that yolov5 outperformed everyone else. I just asked Glenn, the author of yolov5 to drop by and say hi to you.
PS: I read your blog and this post on wheat maturity detection was cool too. Keep creating these projects, many people will find them useful
@lavanyashukla Thank you!
I’m presently getting some experience with wandb, so I hope to be able to create a few blogs that showcase the awesome tool!
Nice, @justintenuto can help you get published on Fully Connected when you have those blog posts!
@maria_rodriguez nice work on the blood cell detection medium post!
Class imbalance is a common concern, but in practice nearly all datasets will suffer this to one degree or another. COCO itself has several orders of magnitude imbalance between the most and least common classes (people and toaster I think).
We’ve tried to create specific tools to address imbalance, like weighted image selection during training with the --image-weights flag, but in practice simply training longer seems to be the best solution, as this naturally provides greater exposure to less frequent classes.
Hey @maria_rodriguez , the article was a great read. I really appreciate you open sourcing code (which btw is very readable, a rare feat to achieve in the medical imaging domain IMHO) and the thoroughness of your experimentation .
@glenn-jocher That’s a valuable insight, thank you for shedding some light to the balance concern!
@sauravmaheshkar Thank you Saurav! It’s uplifting to hear that
Justin and I have been discussing publishing a couple of reports. But he hasn’t been replying to my emails for the past 2 weeks – I hope he is well?