Would 20 images not be a bit a too small number? This could be probably found by testing against a larger sample. Have you tried this?
I didn't know about the Yolo algorithm and I have tried to look for an association of YOLO and jet images in the context of hadronic collisions (jet images are a thing in particle physics). Whereas I have found some hits, I was unable to find actual works that use it (only "rival" methods seem to be considered). I have no idea why (either there is a reason or this has never been tried).
Thank you for your comment :)
It's exactly what I thought at first, but I tried anyway because I thought that simple features of the objects I was trying to detect and the clear background could mean that I could achieve my goal with a small dataset. I didn't take the time to test on various sizes of datasets so here we go :
For this, I used the pets dataset available here: https://www.robots.ox.ac.uk/~vgg/data/pets/ ( and here to have it to the correct format https://public.roboflow.com/object-detection/oxford-pets/1 )
I will only work on one breed, there are around 200 images per breed. So I have uploaded another version here with 2 datsets at 200 epochs :
If you want to test it by yourself you can do it here :
https://colab.research.google.com/drive/1dpNYAnuAY6xdH02ooIkzuujiLSvO97Wo?usp=sharing
Here is the result at the end of the training for the big one :
And for the small one (it's much faster to train and the precision is slightly below) :
But when testing the models, the small one was significantly worse, not detecting most of the dogs, as the one trained with the big dataset found most of the dogs.
So yes, the size of the dataset is important but for features like circles in almost controlled environment 20 pictures were sufficient. But as the complexity of the objects and images increases, you will need bigger datasets. And I found that increasing the number of epochs can compensate a little bit for a dataset too small.
Yeah I haven't found any examples of jet images used with YOLO either, maybe the algorithm isn't fit for this kind of detection. Do you know a small dataset (in case I have to convert the boxes) I could try it on ? I only found this one for the moment : https://zenodo.org/record/3602254#.YfPyn_DMJPY
I have unfortunately no time to test it by myself. I am so overwhelmed at the moment.
For a particle physics dataset, I would need to generate those images by myself. Depending on the time you have and how continuously you could inject energy in such a potential project, I can maybe discuss with a colleague and start a real project with you on this matter. Would it be compatible with your studies?
Cheers!
Unfortunately, I have neither the time nor the expertise for such a project, sorry.
Yeah I knew it ;)