false positives about os2d HOT 8 CLOSED

Hi, it looks like in this particular example it's a matter of the threshold for detections. If you set it to 0.59 all the false positives will be gone.
Best,
Anton

from os2d.

yeah but the threshold cannot be generalized for all cases.

from os2d.

Well, selecting the correct threshold is indeed tricky. There is a trade-off between false positives and undetected objects. The existence of very similar classes makes the problem even harder. If you have data with a lot of similar classes fine-tuning on it can potentially improve performance. We trained the model on the best dataset available to us but there is clear room for improvement.

from os2d.

we have a case were we need to distinguish similar images as above, do you have any recommended method/approach for distinguishing between false and actual class (provided there is not much training data), is there a way for us to take advantage of the color differences in these similar images to be able to distinguish them

from os2d.

We noticed that ImageNet-pretrained weights directly (before fine-tuning) are a bit more sensitive to colors. You can try the OS2D V2-init model which uses them directly. Alternatively, you can merge similar classes into one class with several class images and train an additional classifier to distinguish those - this can be possible without a lot of data.
I'm afraid I do not have significantly more to suggest.

from os2d.

thanks for the help.
also i had a small doubt that do the training of the model require only 1 image per class or the annotated aisle images as in the grozi data set

from os2d.

To train the models, for each class, you need 1 template image and several annotated objects of this class (positives). Optionally, you can use several template images (e.g. use annotated positives as templates) - randomly select one in each batch.

from os2d.

Thank you.

from os2d.